This week’s lead story looks at how soaring AI inference costs are forcing some companies into decisions they probably wouldn't have entertained a year ago — and why one startup's switch from Anthropic to DeepSeek says a lot about where open-weight models now stand.

Elsewhere, Euro-Office launched and immediately drew fire from LibreOffice; a developer built an open source app to liberate Oura Ring data from its subscription paywall; and more.

As usual, feel free to reach out to me with any questions, tips, corrections, or suggestions: forkable[at]pm.me.

Paul

<Open issue>

When the inference bill exceeds payroll, something has to give

Lindy, an AI agent platform that automates everyday work tasks such as email triage, has switched its entire model infrastructure from Anthropic to DeepSeek v4, the open-weight model from Chinese AI research company DeepSeek. The move, announced last week by founder and CEO Flo Crivello, came down to one thing: his AI bill had grown bigger than his wage bill.

“Pulled the trigger today and switched 100% of Lindy traffic to DeepSeek v4, churning from Anthropic models,” Crivello wrote on X. “Saves us millions of $ and we’re actually seeing an increase in performance on many core use cases. Transformative for the business.”

Lindy isn’t alone in hitting this wall. Uber exhausted its entire 2026 AI budget in four months, largely on Claude Code. GitHub scrapped its flat-rate Copilot subscription after agentic coding sessions made the numbers unworkable. The Linux Foundation’s newly launched Tokenomics Foundation — backed by Google, Microsoft, IBM, and Salesforce — exists precisely because no agreed standard exists for enterprises to track or benchmark what AI is actually costing them.

What makes Lindy’s switch significant, from Forkable’s perspective at least, isn’t the saving — it’s what it says about where open-weight models now sit. DeepSeek first rattled the industry in January 2025, when its R1 model went toe-to-toe with leading US frontier models at a fraction of the price. V4, released in preview in April, went further — notable for being the first frontier-class AI stack built entirely within China’s own technology supply chain.

The traffic data tells its own story. Vercel’s AI Gateway recorded DeepSeek’s share of token volume climbing from under 1% to 17% in a single month in May — while its share of actual spend barely moved from 1%, pointing to just how much cheaper these tokens are to run.

Lindy’s migration took far longer than expected — six to nine months of evaluation, gradual rollout, and significant prompt re-engineering. “100x more work than we thought,” Crivello said: “I wouldn’t be surprised if Anthropic’s next release earned them our business back, but they would need to significantly cut prices.”

Read more: The New Stack

<Patch notes>

Euro-Office launches — and LibreOffice isn't impressed

The Document Foundation, the organisation behind LibreOffice, has publicly accused Euro-Office — the European open-source office suite that launched last week as an alternative to Microsoft 365 and Google Workspace — of undermining the very sovereignty it claims to champion.

Read more: The Register

Google joins Eclipse Foundation to push open source developer tools

Google has become a strategic member of the Eclipse Foundation, deepening its involvement in open source developer tooling with a focus on AI-integrated development environments.

Read more: Google

Microsoft open-sources ASSERT to turn specs into agent evals

Microsoft's Responsible AI team has released ASSERT, an open-source framework that converts plain-language behaviour specifications into full evaluation pipelines for AI agents — automatically generating test scenarios, datasets, and scorecards. The idea is simple: if you've written down what your agent should do, ASSERT turns that into something you can actually run and inspect.

Read more: Microsoft | ASSERT

UK backs open source AI builders with compute and a seat at the table

Speaking at London Tech Week, UK AI Minister Kanishka Narayan announced an Open-Source AI Builder Fund providing over £500,000 worth of compute to help developers move prototypes into production, alongside a mentoring scheme and a new dev board giving ten developers under 30 a direct line into government policy.

Read more: UK Gov | Computing

Linux Foundation launches OpenSharing to standardise AI asset exchange

The Linux Foundation has launched OpenSharing, a vendor-neutral open protocol — contributed by Databricks and building on its Delta Sharing project — for sharing AI models, agent skills, and unstructured data across platforms without proprietary marketplaces or custom integrations.

Read more: Linux Foundation

<Final commit>

Cracked Oura: the open source app that breaks the subscription

An Oura Ring will cost you $400 for the hardware, plus up to $70 a year to actually see your health data. Developer Elmo Ahorinta thought that was a bit rich, so he built Cracked Oura — an open source desktop app that pulls your sleep, readiness, and activity data out of Oura's ecosystem entirely and stores it locally in an SQLite database.

It then generates dashboards he says are "at least as good as the official Oura dashboard."

EU data portability rules require Oura to let users export their own data, so Cracked Oura simply automates that process and does the rest itself. No subscription required.

Read more: Android Authority | GitHub

This week’s lead story looks at the European Commission’s landmark tech sovereignty package — and what it means for open source specifically.

Elsewhere, Euro-Office is set to launch as a European alternative to Microsoft 365 and Google Docs; IBM and Red Hat commit $5 billion to securing the open source supply chain; Supabase raises $500 million at a $10.5 billion valuation; and much more.

As usual, feel free to reach out to me with any questions, tips, corrections, or suggestions: forkable[at]pm.me.

Paul

<Open issue>

Europe bets on open source to cut the cord from US and Chinese tech

On Wednesday, the European Commission (EC) published its European Technological Sovereignty Package — a bundle of legislative proposals covering semiconductors, cloud, AI, and, most notably for Forkable, a full Open Source Strategy.

Europe procures more than 80% of its key digital products, services, and infrastructure from non-EU suppliers — a dependence that reared its head last year when China curtailed semiconductor exports, pushing the European car industry to the brink. Add to that the question of what happens if a US administration decides to lean on American cloud providers, be it to cut off services or hand over sensitive European data.

Henna Virkkunen, the Commission’s vice-president for tech sovereignty, said that the EU cannot allow any foreign government or company to hold what she called the “kill switch possibility.” She pointed specifically to the 2018 US Cloud Act, which gives US federal authorities the power to reach into data held by American providers on foreign soil, as something fundamentally at odds with European rules.

As for the open source component of the package, well, some €2 billion over seven years to fund alternatives to proprietary software; a target of 30 million active users of open source collaboration tools by 2030; and a mandate that the EU Digital Identity Wallet be built on open source.

OpenForum Europe, a Brussels-based open source advocacy group, welcomed it as "a defining moment for open source in Europe" — the culmination of more than two decades of work to have open source treated as a foundation of Europe's digital future.

“This is Europe's most ambitious commitment to open source to date,” said Nicholas Gates, senior policy advisor at OpenForum Europe. “The devil is always in the details, but this is a remarkable step forward for open source in digital policy, and the Commission should be lauded for getting us this far.”

Even within the open source camp, there are doubts about some of the specifics of the sovereignty package: €2 billion spread thinly across seven years looks modest set against the Commission’s own figure of €264 billion spent annually on largely proprietary IT. Open source hardware barely features, too.

On the cloud provisions specifically, the Computer and Communications Industry Association (CCIA) — whose members include Amazon and Google — called the proposals a “dangerous recipe for progressive market shutdown,” arguing the move was “discriminatory.”

“By excluding trusted international technology providers based on their headquarters location and organisational structure, the Commission forces users to rely on a much more limited selection of digital products,” the CCIA wrote.

The proposals still require agreement from EU member states and the European Parliament, so there is a long way for this to get over the line. But Brussels has now, for the first time, placed open source at the centre of how Europe intends to take back control of its own digital infrastructure.

Drupal founder and project lead Dries Buytaert called the new open source strategy “a big step forward.”

“It treats Open Source as infrastructure, ties it to technological sovereignty, and backs it with real funding,” Buytaert wrote.

Read more: European Commission | Tech Policy Press | The Guardian

<Patch notes>

Europe's answer to Microsoft Office to go live

Speaking of sovereignty, the first stable version of Euro-Office is set to launch on June 9, courtesy of a coalition of European companies which have built an open-source office suite designed as a direct alternative to Microsoft 365 and Google Workspace for governments, public sector bodies, and regulated organisations.

Read more: ZDNET | Euro-Office

Tokenomics Foundation to tackle AI cost chaos

AI token costs are rising fast and nobody agrees on how to measure them. With that in mind, the Linux Foundation this week announced the Tokenomics Foundation, a new body tasked with building open standards, benchmarks, and best practices across the AI token economy, backed by Google, Microsoft, IBM, JPMorgan Chase, and Salesforce. Notably absent: Anthropic and OpenAI, the frontier model providers whose pricing sits at the heart of the problem.

Read more: The New Stack

JetBrains open-sources Mellum2

JetBrains has released Mellum2, an open-source coding model built to run entirely on your own infrastructure — a direct counter to coding tools that depend on third-party APIs.

Read more: The New Stack

IBM and Red Hat commit $5 billion to open source security

IBM and Red Hat launched Project Lightwell, an enterprise clearing house for securing the open source software supply chain, backed by a $5 billion commitment.

Read more: Red Hat

Netflix engineer open-sources AI cost-cutting tool

A senior Netflix engineer has released Project Headroom, an open-source tool that promises to cut LLM costs by stripping redundant tokens before they reach a model. Since its January release, the project has reportedly saved users around $700,000 and freed up 200 billion tokens.

Read more: The Register

Rayfin lands to bridge vibe coding and enterprise production

Microsoft has released Rayfin, an open-source SDK built to take apps created with AI-assisted vibe coding tools and make them fit for enterprise deployment, with security and governance baked in.

Read more: Microsoft

Cloudflare acquires VoidZero

Cloudflare has acquired VoidZero, the team behind Vite, Vitest, Rolldown, and Oxc — some of the most widely used open-source JavaScript build tools in the ecosystem. Cloudflare says Vite will remain open source and vendor-agnostic.

Read more: Cloudflare

Supabase raises $500 million at $10.5 billion valuation

Open-source Postgres development platform Supabase (previous coverage here) has closed a $500 million series F round of funding, doubling its valuation to $10.5 billion.

Read more: CNBC

Yugabyte launches Meko, a memory layer for AI agents

Open-source database company Yugabyte launched Meko, a shared memory layer designed to give AI agents persistent, consistent context across tasks and sessions.

Read more: Forkable

OpenTelemetry hits general availability

Seven years in the making, OpenTelemetry — the open-source observability framework that has become the telemetry standard for cloud computing — formally reached general availability. The project is now also positioning itself for the AI infrastructure era.

Read more: The New Stack

Gas Town comes to the cloud

Gas Town, the open-source AI agent orchestration project created by developer and writer Steve Yegge, has received a cloud-hosted version via Kilo. The new Wasteland feature supports running a thousand Gas Town instances simultaneously.

Read more: The New Stack

Ex-Block CTO goes all in on open agentic AI infrastructure

Manik Surtani has left his role as head of open source at payments company Block after eight years, to become CTO of the Agentic AI Foundation, a new organisation focused on keeping the next layer of AI infrastructure open.

Read more: Substack

<Final commit>

Open skies

Researchers at MIT and the University of Pennsylvania have published MIGHTY, an open-source trajectory planning system for autonomous drones that can avoid unexpected obstacles in real time while keeping to a smooth, time-efficient flight path.

The system runs entirely on the drone’s own onboard hardware, negating the need for the commercial solvers that currently underpin most high-performance alternatives — software that can carry price tags running into the hundreds of thousands of dollars. In tests, MIGHTY cut computation time by 9.3% and travel time by 13.1% compared to existing approaches, with a 100% success rate. Potential applications include search-and-rescue in collapsed buildings, last-mile delivery in cities, and infrastructure inspection.

The code is available on GitHub.

Read more: MIT News | Mighty (GitHub)

In this edition, I check in with Karthik Ranganathan, CEO and co-founder of Yugabyte, an open-source database company now pushing into AI infrastructure with the launch of Meko, a new “memory layer” for AI agents.

Meko reflects how infrastructure companies are reassessing their role in the agentic AI stack. As agents take on longer-running tasks across multiple systems, the problems companies keep hitting — shared context and coordination — are ones distributed systems engineers would recognise immediately.

"Agents are working from inconsistent views of what's true right now," Ranganathan told me. "This is because there is no way to enforce a shared truth."

Read the story in full below.

10 years in the making

When Karthik Ranganathan launched open-source database startup Yugabyte some 10 years ago, the company was focused on helping enterprises run PostgreSQL applications reliably — essentially, making sure apps and services stayed online and responsive even when spread across different data centers around the world.

AI agents, and the strain they would eventually place on data systems, weren’t part of the conversation.

Ranganathan arrived at Yugabyte after years building distributed database technology at Facebook, where he worked on projects including Apache Cassandra and HBase, before later joining cloud infrastructure company Nutanix. Yugabyte itself grew out of the cloud computing boom.

Now the company is turning its attention squarely towards AI agents. Last week, Yugabyte launched Meko, a “memory layer” designed to help AI agents store, share, and retain knowledge across tasks and between different agents.

The release comes as companies building agentic systems run into mounting difficulties around maintaining shared context, keeping information current, and tracking what agents have learned over time.

Many companies currently piece together agent memory using separate retrieval systems, databases, logging tools, and orchestration software. According to Ranganathan, that can create situations where different agents end up operating with different versions of the same information.

“Agents are working from inconsistent views of what's true right now,” Ranganathan explained to Forkable. “This is because there is no way to enforce a shared truth.”

As different agents interact with the same systems, one may update a piece of information while another continues working from an older version, causing errors and outdated assumptions to spread between agents over time.

“There’s no consistency model anywhere in a stitched stack,” Ranganathan continued.

Meko and the next layer of AI infrastructure

Meko, at its core, is a shared memory layer for AI agents. The software stores things like conversation history, agent decisions, shared knowledge, and longer-term memory in a single system that different agents can access. Yugabyte says the goal is to help agents retain context across tasks and sessions, while also giving teams a clearer record of how information moves between systems.

“Existing tools trace what agents did — they don't trace what agents learned, or how that learning propagated between agents,” Ranganathan said.

At the center of the platform are four main components: memory, knowledge, conversations, and traces. Together, those pieces are intended to handle persistent agent memory, shared organisational knowledge, conversational history, and records showing how agents arrived at particular decisions over time.

Ranganathan argues that many current AI tools can show prompts, tool calls, and latency, but struggle to explain how agents arrived at certain conclusions or how information spread between different systems over time. He also believes that growing regulatory pressure around AI accountability, including requirements around auditability, will place more scrutiny on those gaps as agentic systems take deeper root in enterprise software.

“None of these are fixable with better prompts or more retrieval,” Ranganathan continued. “They're infrastructure problems that need infrastructure solutions.”

The open source factor

For now, Meko is available by request only as a fully managed service, though Yugabyte says the intention is to release the underlying platform as open source. And the rationale for that, according to Ranganathan, is largely driven by the nature of the technology itself, and the belief that companies will be reluctant to build critical AI systems on top of closed memory infrastructure they cannot independently inspect or control.

“Memory infrastructure for agents will be deeply embedded in production applications — it holds the learned context, the conversation history, the audit trail of what agents knew and when,” Ranganathan explained. “That layer can’t be a proprietary black box. If we asked teams to commit their agent data to a closed system that they can’t self-host, audit, or extend, we’d lose them on day one. The data is too important to trust to a single vendor.”

The launch also reflects Yugabyte’s longer-running business model around open source infrastructure. The company has historically monetised through managed services built on top of its open source database technology.

Meko will ultimately follow a similar pattern, with Yugabyte making money through hosted and managed deployments aimed at enterprises that don’t want to operate the infrastructure themselves. Underneath, the system runs on YugabyteDB, the company’s PostgreSQL-compatible distributed database, which acts as the storage and coordination layer behind the memory system.

For Yugabyte, it’s also a bet on where enterprise software is heading — databases and data platforms angling for a foundational role in agentic systems, without the lock-in.

“If you’re building a multi-agent system where models will change, vendors will change, and you need an audit trail across the whole thing — that’s an infrastructure decision,” Ranganathan said. “And it shouldn’t be locked to whichever model you happened to start with.”

This week’s lead story looks at NHS England moving hundreds of public GitHub repositories behind closed doors, amid growing concern that AI systems capable of analysing codebases at scale are changing the security calculus around open source.

Elsewhere, China’s open-model boom continues with another massive funding round, AWS and Cisco unveil new infrastructure for securing AI agents and model supply chains, and a fresh wave of startups and investors pile into the tooling, inference, and orchestration layers forming around open AI systems.

As usual, feel free to reach out to me with any questions, tips, corrections, or suggestions: forkable[at]pm.me.

Paul

<Open issue>

NHS locks down public code

AI-driven vulnerability discovery is starting to reshape how public institutions think about open source.

England’s National Health Service (NHS), one of the world’s largest public healthcare systems and among the biggest employers globally, has instructed teams to make hundreds of public GitHub repositories private by May 11, amid growing concern that modern AI systems can analyse open codebases for weaknesses at scale.

According to internal guidance seen by The Register, the move is tied directly to advances in AI-assisted security research, including Anthropic’s Mythos project, which recently demonstrated how frontier models can identify vulnerabilities in widely used software that had survived years of human review.

Notably, NHS isn’t responding to a known breach. Instead, it appears to be reacting to a broader concern: that the economics of vulnerability discovery have changed. Systems capable of ingesting and reasoning over large codebases may make it significantly easier — and cheaper — to identify weaknesses in publicly accessible software and infrastructure. By way of example, Anthropic said that Mythos had found a flaw in OpenBSD, a security-focused operating system, dating back 27 years.

The repositories affected reportedly include a broad mix of projects, from software tools and scripts to documentation and architecture resources. Many were public because NHS England, like other parts of the UK public sector, had increasingly embraced open-source principles around transparency, reuse, and reducing duplication across the public sector.

That context makes the move notable. For years, government-backed open-source efforts have operated on the assumption that openness improves software quality and security by allowing more people to inspect, test, and improve code. The NHS decision suggests some organisations are beginning to worry that the same visibility may now also lower the barrier for attackers equipped with increasingly capable AI tools.

The move also lands amid a broader debate unfolding across open source. In recent weeks, companies including Cal.com have argued that AI-assisted vulnerability research is forcing them to rethink how much code they expose publicly.

But the NHS case could be more consequential because it extends beyond startups and into national infrastructure. Open repositories are often used not just for software distribution, but for collaboration between agencies, procurement, and public accountability around taxpayer-funded systems.

For now, NHS says the move is temporary while it reassesses risk and governance around public code. Whether that becomes a lasting policy change may depend on how seriously institutions come to view AI-assisted vulnerability discovery — and whether openness itself is increasingly seen as part of the attack surface.

Read more: The Register

<Patch notes>

Moonshot AI lands $2B as China’s open-model boom grows

Chinese AI lab Moonshot AI has raised $2 billion at a $20 billion valuation, amid surging demand for open-weight AI models. The company’s Kimi models are gaining traction as China continues pushing an increasingly open AI ecosystem.

Read more: TechCrunch

New fund targets open-source science software

A new Open Source for Science Fund has launched to support the open software stack underpinning scientific research.

Read more: Open Source for Science Fund | OS 4 Science

GitHub adds security scanning for MCP agents

GitHub has launched dependency and secret scanning for its MCP server, allowing AI coding agents to catch leaked credentials and vulnerable packages before code is committed.

Read more: The New Stack

Anthropic donates open-source Petri toolkit

Anthropic is donating Petri, its open-source AI alignment and interpretability toolkit, to researchers and nonprofits working on model transparency and safety. Its new home is at Meridian Labs, an AI evaluation nonprofit.

Read more: Anthropic

Cisco launches model provenance toolkit

Cisco has released Model Provenance Kit, an open-source framework for tracking how AI models are built, modified, and distributed — part of a broader push toward traceability in AI supply chains.

Read more: Cisco

RadixArk launches with $100M

New startup RadixArk has launched with $100 million in seed funding to expand SGLang, the fast-growing open-source inference framework for large language models. The company says it wants to “democratize frontier AI infrastructure.”

Read more: Business Wire

CopilotKit raises $27M for app-native AI agents

Open-source AI agent framework CopilotKit has raised $27 million to help developers build app-native agents directly into their own software, rather than relying on standalone chatbots and hosted assistants.

Read more: TechCrunch

DeepInfra raises $107M for open-model inference cloud

DeepInfra has raised $107 million to build dedicated cloud infrastructure for serving open-source AI models, betting demand for inference outside the hyperscalers will continue growing.

Read more: Silicon Angle

<Final commit>

A healthy approach to wearable data…

Open Wearables, an open-source health intelligence platform built by healthtech studio Momentum, has released open-source scoring algorithms for wearable data, starting with sleep and resilience metrics typically locked behind platforms like Whoop and Oura.

The system connects data sources including Strava, Garmin, Apple Health, and Google Health Connect through a unified API, while also providing an AI reasoning layer for analysing trends and health signals. The broader pitch is that health scoring — and increasingly AI interpretation of wearable data — should be transparent and auditable.

Read more: Fitt/Insider

This week’s lead story looks at a widely used open-source package being compromised through its own release process, after attackers slipped a malicious update into a trusted distribution channel.

Elsewhere, a new crop of tools and funding rounds points to growing momentum around AI agents and the infrastructure that supports them, while governments and developers continue to rethink their reliance on dominant platforms — from sovereign Git hosting in Europe to high-profile departures from GitHub itself.

As usual, feel free to reach out to me with any questions, tips, corrections, or suggestions: forkable[at]pm.me.

Paul

<Open issue>

Hackers weaponize trusted update

A widely used open-source package with more than 1 million monthly downloads was compromised after attackers hijacked its release process and pushed a malicious update.

The package, elementary-data, was briefly distributed with credential-stealing code that searched systems for sensitive data including cloud keys, API tokens, SSH keys, and warehouse credentials. The malicious version — 0.23.3 — was live for around 12 hours before being pulled.

The attack didn’t rely on tricking users into installing something obviously suspicious. Instead, the attacker exploited a vulnerability in a GitHub Actions workflow used by the project’s maintainers. By submitting a malicious pull request, they were able to execute code inside the developers’ environment, extract signing keys and tokens, and publish a compromised release that appeared legitimate.

“Users who installed 0.23.3… should assume that any credentials accessible to the environment where it ran may have been exposed,” the Elementary team warned.

The incident follows a well-trodden path. Rather than targeting the codebase itself, attackers are going after the systems that build and ship it. In this case, the release pipeline — the automated process that builds the software, signs it, and publishes updates — became the point of entry.

For maintainers, that creates a difficult problem. Code is reviewed and tested, but CI pipelines — often less scrutinised and holding the keys to production — are easier to get wrong and far more damaging if compromised.

Read more: Ars Technica | Elementary

<Patch notes>

Paper Compute wants to fix AI agent infrastructure

GitHub veteran Brian Douglas recently launched Paper Compute, an open-source startup building infrastructure for AI agents, starting with Tapes (observability) and StereOS (a hardened OS for sandboxed agent execution).

Read my interview with Brian in the link below.

Read more: The New Stack

Netherlands launches ‘sovereign’ GitHub alternative

The Dutch government has soft-launched code.overheid.nl, a self-hosted Git platform for public bodies built on Forgejo, a European alternative to GitHub.

Read more: Digital Government | Code.Overheid

Hashicorp co-founder quits GitHub

Mitchell Hashimoto is moving his Ghostty terminal project off GitHub after near-daily outages. "GitHub is no longer a place for serious work," he wrote, after 18 years on the platform.

Read more: The Register

Warp goes open source

AI coding startup Warp is open-sourcing its client, betting a community managing AI agents can outpace well-funded closed rivals. OpenAI is a founding sponsor.

Read more: Tessl

OpenAI open-sources Symphony

OpenAI this week released Symphony, an open source specification for orchestrating coding agents. The project connects issue trackers such as Linear to Codex-powered agents, allowing each task to run as its own self-contained job, from assignment through to a pull request.

Read more: Tessl

Vendor lock-in drives OSS adoption?

The 2026 State of Open Source Report is out! Vendor lock-in is now cited by 55% of organisations as a primary driver of OSS adoption — a 68% year-over-year jump — while security and maintenance remain the biggest operational challenges.

Read more: The Open Source Initiative (OSI) | Perforce Open Logic

OpenAI releases local PII filter

OpenAI’s new Privacy Filter model detects and redacts personally identifiable information (PII) on-device, supporting up to 128k tokens in a single pass.

Read more: The New Stack | OpenAI | Hugging Face

Cursor partners with Chainguard to secure AI-built code

Cursor is routing dependencies through Chainguard's verified package repositories, addressing the risk of agents pulling compromised libraries at a pace no security team can review manually.

Read more: Tessl

Intel shutters its open-source evangelism program

As part of its ongoing “open-source rethink,” Intel has archived its Open Ecosystem Community/Evangelism initiative alongside a wave of GitHub repositories, marking a significant retreat from its decades-long role as a major open-source contributor.

Read more: Phoronix

DeepSeek V4 arrives with Huawei chip support

DeepSeek's most significant release since R1 comes in Pro and Flash variants with a 1M token context window, and is the first model optimised for domestic Chinese chips.

Read more: MIT Technology Review

Cloudsmith raises $72M

Belfast-based artifact management platform Cloudsmith has raised a $72 million Series C led by TCV and Insight Partners, targeting enterprises managing AI-generated code at scale.

Read more: Cloudsmith

ComfyUI raises $30M at $500M valuation

Open-source AI workflow platform ComfyUI has raised $30 million at a $500 million valuation. The company now claims 4 million users and 150,000 daily downloads.

Read more: GlobeNewswire

<Final commit>

Microsoft digs up the bones of DOS

On the 45th anniversary of 86-DOS 1.00, Microsoft has published what appears to be the earliest-known DOS source code — hand-annotated assembler printouts preserved by Tim Paterson, the original author of DOS. Think scanned listings, handwritten notes, and the kind of artifacts that you might find in an archaeological dig.

Microsoft describes them as “a printed commit history of a Git repository” — a stack of paper that tracks which features were added when, what broke, and how it got fixed. The materials include the 86-DOS 1.00 kernel, development snapshots of PC-DOS 1.00, and utilities like CHKDSK. Soon, the physical originals will be on display at the Interim Computer Museum, donated by Paterson himself.

Read more: Microsoft

This week’s lead story looks at Cal.com moving its core codebase behind closed doors, as the company argues that AI-driven vulnerability discovery is making open development harder to defend.

Elsewhere, Expo hires Meta’s former React lead as it raises fresh funding and launches a new agent for building mobile apps; France pushes ahead with plans to move public systems from Windows to Linux; and a mix of security issues, new tools, and governance updates continue to shape the open-source ecosystem.

As usual, feel free to reach out to me with any questions, tips, corrections, or suggestions: forkable[at]pm.me.

Paul

<Open issue>

Cal.com draws a line under open source

AI is putting new pressure on open source. Maintainers are already battling a surge of machine-generated “AI slop,” but the bigger challenge might be just beginning: tools that can scan codebases and surface vulnerabilities at scale.

Last week I reported on Anthropic’s Claude Mythos and Project Glasswing, and the rise of AI systems that can hunt for vulnerabilities in widely used software. This week, Cal.com, an open-source Calendly alternative that provides scheduling infrastructure for developers, is pushing its core codebase behind closed doors.

Cal.com has moved its main production codebase into a private repository, while releasing a stripped-down, MIT-licensed community version called Cal.diy for self-hosting.

CEO Bailey Pumfleet argues that openness now comes with a cost. He said access to source code lowers the barrier to finding weaknesses. “It’s easier to perform a bank heist if you have the blueprints to the vault,” Pumfleet said. “It’s a lot easier to see the inner workings of something and reverse engineer it to find a vulnerability.”

Asked whether stronger hardening could offset the risk, Pumfleet said: “Simply having the code open increases the risk dramatically.

“It’s not closing the code or hardening it; it’s just that we’re doing both,” he continued. “Simply hardening it does not decrease the risk enough.”

The argument arrives as tools like Mythos have shown they can uncover vulnerabilities that sat undetected for decades, including a flaw in OpenBSD dating back 27 years.

Not everyone accepts this logic, however. Critics argue that closing code reduces the number of people who can find and fix bugs, and that security depends on scrutiny rather than secrecy. Others suggest the move reflects commercial pressure dressed up as a security decision.

Developer Simon Willison pointed to recent analysis by Drew Breunig, which argues that AI-driven vulnerability discovery raises the cost of securing software for everyone — and that open source helps spread that burden across a wider pool of contributors, rather than concentrating it within a single company.

Pumfleet, for his part, agrees with this view, but said that it all comes down to a question of resources. He said approaches that rely on continuous, large-scale analysis would require budgets Cal.com doesn’t have, adding that closing the codebase is “the biggest method of risk reduction we can take right now to secure our customers.”

There is also a longer history behind such actions. Companies built on open source have frequently restricted access to parts of their codebases — whether through licensing changes, enterprise carve-outs, or fully closing core components. That context has led some critics to question whether security is the sole driver here, or whether it also serves as a justification for tighter control.

Pumfleet rejects that framing. “The decision is entirely about security,” he said. “We already have control over the product being open source, and we don’t really stand to gain much from being closed source. It’s really just a question about de-risking on the security side.”

Read more: The New Stack | Cal.com | Drew Breunig

<Patch notes>

Expo hires Meta’s React lead, launches agent, nabs $45M

Open-source React Native development platform Expo has hired Meta’s former React lead Seth Webster, who most recently emerged as executive director of the all-new React Foundation after Meta donated the project.

In tandem, Expo also announced $45 million in funding and the launch of Expo Agent. Expo sits on top of React Native, handling much of the infrastructure needed to ship apps, and it’s now pushing into agent-driven development by combining code generation with deployment and testing.

Read my interview with Seth Webster for The New Stack on the link below.

Read more: The New Stack

France ditches Windows for Linux

In yet another sovereign tech push from Europe, France is moving public sector systems off Windows and onto Linux to reduce reliance on US tech vendors.

Read more: TechCrunch

EU age-checking app already cracked

The EU is rolling out an open-source age verification app, but hackers say they broke it in minutes. It raises questions about whether these systems can actually protect kids—or just create new risks.

Read more: European Commission | Politico

Backdoors found in WordPress plugins

Dozens of WordPress plugins used by thousands of sites were found with hidden backdoors.

Read more: TechCrunch

Microsoft shares Surface data with Linux Foundation

Microsoft is handing over key Surface device data to the Linux Foundation’s Battery Data Alliance to help improve battery life on Linux.

Read more: TechRadar

Nvidia launches quantum-focused AI models

Nvidia unveiled new open models aimed at accelerating quantum computing research.

Read more: Tom’s Hardware | Nvidia

Mozilla launches Thunderbolt AI client

Mozilla has launched a new AI client called Thunderbolt, focused on running models locally or in self-hosted setups. The pitch is tighter control over data and infrastructure, aimed at developers and teams who don’t want to rely on hosted AI services.

Read more: Ars Technica | Thunderbolt

Vercel open-sources AI agent framework

Vercel released Open Agents, a toolkit for building custom AI coding agents. The idea is to let companies run their own agents on their own infrastructure, rather than relying on hosted tools.

Read more: Tessl

Laude’s moonshots

Laude Institute, a nonprofit funding critical open research, has unveiled eight academic teams as finalists in its new “Moonshots” program. The groups are working in the open and will compete over the next six months for $10M lab-scale funding to help solve “humanity’s hardest problems.”

Read more: Laude Institute

OSI names new executive director

The Open Source Initiative (OSI) has appointed Duane O'Brien as executive director, some six months after his predecessor Stefano Maffulli announced he was stepping down.

Read more: The Open Source Initiative

<Final commit>

Linux allows AI! But not the sloppy kind…

Linux 7 landed this week, bringing the usual mix of hardware support updates, performance tweaks, and under-the-hood changes that keep the kernel ticking.

However, arguably the more interesting facet was a little tidbit that came alongside it: a new policy on AI-assisted contributions to the kernel. AI tools are officially allowed, but with important caveats.

AI agents mustn’t add “Signed-off-by” tags, with only humans able to legally certify the Developer Certificate of Origin (DCO).

As per official documentation, the human submitter is responsible for:

• Reviewing all AI-generated code

• Ensuring compliance with licensing requirements

• Adding their own Signed-off-by tag to certify the DCO

• Taking full responsibility for the contribution

Put simply, AI is allowed, but no slop.

Read more: The Register | GitHub

This week’s lead story looks at Project Glasswing, an effort to bring frontier AI security tools into the hands of open-source maintainers (though not exclusively), as models like Anthropic’s Claude Mythos begin to uncover vulnerabilities that have gone undetected for decades.

Elsewhere, there’s a continued push toward open messaging infrastructure in Europe, new tools to secure AI agents, and a mix of projects and platforms shaping how open infrastructure is built and governed, and more.

As usual, feel free to reach out to me with any questions, tips, corrections, or suggestions: forkable[at]pm.me.

Paul

<Open issue>

Glasswing brings frontier AI to open source security

Anthropic says its shiny new, cutting-edge Claude Mythos model has reached a level of coding and security capability where it can find and exploit software vulnerabilities that survived years — and in some cases decades — of human review.

While the ramifications of this are far-reaching, Mythos is a big deal for the open source world in particular. Through Project Glasswing, the Linux Foundation, Anthropic, and a group of large tech companies including Amazon, Apple, Google, and Microsoft are putting Mythos to work on defensive security across critical software systems, including widely used open-source projects.

Anthropic is committing up to $100 million in usage credits to the effort, alongside $2.5 million for Alpha-Omega and OpenSSF through the Linux Foundation, plus $1.5 million for the Apache Software Foundation.

“Open source security has historically been a thankless task,” The Linux Foundation’s CEO Jim Zemlin wrote. “Triaging and fixing bugs, writing and testing patches, crafting careful communications strategies – none of this is what maintainers had in mind when they sent their first project commit. We believe AI can help address this.”

Open source underpins much of the modern software stack, but many maintainers are already overwhelmed by bug reports, supply chain attacks, and AI-generated noise. Now they may also have to contend with AI systems that are exceptionally good at surfacing zero-days.

To underscore this point, Anthropic published a handful of examples. Mythos, it said, found a flaw in OpenBSD, a security-focused operating system, dating back 27 years.

In short, Glasswing isn’t just about AI helping create patches a bit faster. It’s about getting frontier defensive capability into the hands of maintainers before the same class of models is used more aggressively elsewhere.

Zemlin stressed that access — not just capability — will determine whether these tools actually help the open-source ecosystem.

“Because the dark side of AI-augmented security is AI-augmented insecurity, we must ensure that access to the best AI cybersecurity tooling is evenly distributed and not concentrated in the hands of the few with the cash and the headcount,” Zemlin noted. “None of this matters if the cost is prohibitive. Project Glasswing is designed to ensure that maintainers get access to these tools for free. This is the only way to foster wide adoption of top AI cybersecurity capabilities – by removing any economic friction.”

Read more: The New Stack & The Linux Foundation & Glasswing & Mythos Preview

<Patch notes>

A ‘beam of light’ for EU sovereignty

Belgium is rolling out a Matrix-based messaging app for government use, in a push toward digital sovereignty. The move leans on open protocols to reduce reliance on proprietary communication platforms.

Read more: Resilience Media

Microsoft open-sources agent security toolkit

Microsoft released the Agent Governance Toolkit, an open-source runtime security layer for AI agents. The project focuses on monitoring behaviour and enforcing guardrails as agents take on more tasks autonomously.

Read more: Microsoft & GitHub

A ‘spiritual successor to WordPress’

Cloudflare has unveiled Emdash, a new open-source CMS positioned as a modern, serverless successor to WordPress. WordPress co-creator Matt Mullenweg said the project was likely designed to drive usage of Cloudflare’s own services, though overall he was complementary of the project.

Read more: Cloudflare & GitHub & Matt Mullenweg

Google releases Gemma 4

Google has launched Gemma 4, the latest version of its open model family aimed at developers. The release continues Google’s push to provide smaller, more accessible models alongside its closed systems.

Read more: Google

Netflix goes into the VOID

Netflix has released VOID, an open-source model that removes objects from video — including their effects on the scene, such as shadows, reflections, and physical interactions. The model builds on CogVideoX and is fine-tuned for interaction-aware video editing in post-production.

Read more: Forbes & VOID (GitHub)

Meta shifts on open models

Meta this week introduced Muse Spark, its first frontier model and the first without open weights, marking a break from its earlier approach. The company says it still plans to release some models openly, though questions remain over how far that commitment will go.

Read more: Meta & Axios & The New Stack

Microsoft deactivates dev accounts

Microsoft deactivated two notable open source developer accounts as part of a new developer verification program, raising concerns about access and control over software distribution platforms. The move highlights ongoing tension between platform governance and open development.

Read more: The Register

<Final commit>

A DIY retro camera

An open-source camera project called Saturnix is gaining a little steam, built around a Raspberry Pi with a 16MP sensor and styled after point-and-shoot cameras from the 1990s.

It’s not a finished product you can buy. Instead, the creator has published the design files, code, and instructions, so users assemble it themselves using off-the-shelf parts like a Raspberry Pi, camera module, and custom 3D-printed case.

The appeal is in that setup: the software can be modified, and the hardware can be rebuilt or adapted as needed. It’s still early, but it offers a glimpse of what open hardware can look like when the full stack — from casing to code — is available to the user.

Read more: Hackaday & GitHub

This week’s lead story looks at the launch of Betterleaks, a new open-source secrets scanner from the creator of Gitleaks, and what it says about how security tooling — and ownership — is evolving in the AI era.

Elsewhere, China is pulling ahead on open AI models much to the chagrin of the US; Google expands Android Automotive deeper into the car stack; Cisco rolls out a new open-source tool to secure AI systems; and a fresh round of donations and project graduations points to continued momentum in open infrastructure and governance.

As usual, feel free to reach out to me with any questions, tips, corrections, or suggestions: forkable[at]pm.me.

Paul

<Open issue>

Secrets scanning for the agentic era

Gitleaks is one of the most popular open-source secret scanning tools, ratcheting up more than 25,000 GitHub stars and tens of millions of downloads since its launch in 2018.

Now, Gitleaks creator Zach Rice is starting afresh with Betterleaks, a new open-source scanner designed to catch exposed credentials — API keys, tokens, passwords — as code is written and shared.

So… why launch a new thing when the old thing was seemingly doing so well? The answer traces back to Rice’s move to Truffle Security in 2023, where his focus shifted to rival open-source scanner TruffleHog. Development on Gitleaks slowed, he said, and Rice no longer had full control over the project.

Betterleaks is positioned as a continuation — a drop-in replacement, with existing commands and configurations still working, but with a reworked detection approach, faster scanning, and more flexibility over how secrets are identified and validated.

Rice says the goal is to keep moving development forward without forcing the existing community to migrate.

“Hopefully it’s not going to cause too much of a backlash to the community – I love the Gitleaks community, and I don’t want to fracture that,” Rice told me in an interview for The New Stack this week. “So if you want to continue using Gitleaks, feel free. It’s stable — and security patches and stuff like that, I’ll continue to do. But if you want the next generation of Gitleaks and the evolution, then switch to Betterleaks.”

Betterleaks is also being built with AI-assisted development in mind. As more code is generated quickly — often with less review — the risk of accidentally exposing credentials increases.

The project is sponsored by Aikido, a billion-dollar security startup that’s backing a broad array of open source security tooling, including OpenGrep, Zen, and Safe Chain.

Betterleaks, for its part, is available under an MIT license, with Rice retaining ownership. It’s also worth noting that there are no plans to build a commercial entity off the back of Betterleaks — the focus, instead, is on building and maintaining a tool that others can rely on, even competitors.

“Like what Aikido did with OpenGrep (forked Semgrep) we’re dedicated to providing really great open source projects for the security community,” Rice said. “A strong open source project is the backbone of a lot of the security products out there. Yes, it’s beneficial to other companies, but it’s also really beneficial to Aikido to have these stable projects.”

Read more: The New Stack

<Patch notes>

China leads on open AI models

As per a recent report from Hugging Face, Chinese AI labs are leading on open AI models — though not the infrastructure behind them (that’s Nvidia) — with data showing they dominate downloads and usage globally. In turn, this is raising concerns among US policymakers that China’s warm embrace of open AI models could erode the US’s lead in AI development.

Read more: The New Stack & Hugging Face & SCMP

Android Automotive moves beyond infotainment

Google is expanding its Android Automotive OS to handle more of the car’s internal systems beyond infotainment, as vehicles shift toward software-defined architectures.

Read more: The Verge

Cisco launches DefenseClaw

Cisco this week introduced DefenseClaw, an open-source tool for monitoring AI agents and detecting issues like prompt injection and unsafe tool.

Read more: Cisco

Fivetran donates SQLMesh to Linux Foundation

Data integration giant Fivetran has donated SQLMesh, its open-source data transformation framework, to the Linux Foundation.

Read more: The New Stack

Tekton and Fluid move up at CNCF

Tekton, a Kubernetes-native CI/CD framework, and Fluid, a data orchestration system for AI and big data workloads, have both been promoted to incubating status by the Cloud Native Computing Foundation (CNCF).

Read more: CNCF and CNCF

<Final commit>

Built on open, sold as new

AI coding company Cursor last week launched Composer 2, an AI model designed to handle longer programming tasks at lower cost.

However, arguably the more interesting detail is that it’s built on Kimi K2.5, an open model from Chinese AI company Moonshot AI. This helps demonstrate how smaller players can build competitive products without training frontier systems like those from OpenAI, Anthropic, or Google.

That detail only emerged after developers dug into the system, prompting criticism over how the underlying model was credited (or not) — a long-running tension in the open source sphere.

It’s a reminder that open models are changing who is able to compete, while raising fresh questions about who gets the credit.

Read more: Tessl & VentureBeat

This week’s lead story looks at OpenAI’s move to acquire Astral, the company behind popular open-source Python tools Ruff and uv, as it continues to pull widely used developer tools and viral projects into its orbit.

Elsewhere, there’s fresh funding for open-source security, growing corporate interest around OpenClaw, and a lot more.

As usual, feel free to reach out to me with any questions, tips, corrections, or suggestions: forkable[at]pm.me.

Paul

<Open issue>

OpenAI’s Astral projection

OpenAI this week announced plans to acquire Astral, the company behind widely used open-source Python tools including Ruff and uv. The plan is to fold Astral into OpenAI’s Codex team, which builds the company’s AI coding agent, and connect Codex “more directly with the tools developers already use.”

Astral has built a following by replacing slower, fragmented Python tooling with faster, unified alternatives. Ruff acts as a high-speed linter and formatter, while uv rethinks package management and environments — both designed to speed up day-to-day development.

The deal gives OpenAI a foothold in the tooling developers actually use when writing and running code, far beyond the model layer.

The move follows OpenAI bringing in the creator of OpenClaw, the viral open-source agent that spread quickly as developers forked and adapted it. It also comes alongside a separate deal for Promptfoo, an open-source testing tool.

While OpenAI mostly abandoned its open-source roots a long time ago, these recent deals are a tacit acknowledgement of the power of open source — and the communities that drive it.

Astral’s tools have grown rapidly these past few years, reaching hundreds of millions of downloads per month and becoming a mainstay of modern Python development. Founder and CEO Charlie Marsh said that growth exceeded expectations.

“Open source is at the heart of that impact and the heart of that story; it sits at the center of everything we do,” Marsh wrote. “In line with our philosophy…. OpenAI will continue supporting our open source tools after the deal closes. We'll keep building in the open, alongside our community – and for the broader Python ecosystem – just as we have from the start.”

Read more: OpenAI & Astral

<Patch notes>

$12.5M for open source security

The Linux Foundation has secured $12.5 million in new funding from major tech firms — including Anthropic, AWS, GitHub, Google, Google DeepMind, Microsoft, and OpenAI — to support open-source security projects. The effort targets long-standing gaps in maintaining widely used but under-resourced infrastructure.

Read more: Press release

Germany backs open document standards

Germany’s sovereign digital stack will mandate the Open Document Format (ODF) — an open, standardised file format used by office suites like LibreOffice — reinforcing a shift away from proprietary file formats in public sector systems. The move aims to improve interoperability and reduce vendor lock-in.

Read more: The Document Foundation

MCP maps next steps

The folks behind the Model Context Protocol (MCP) have outlined a set of priorities for 2026, including better support for remote servers, improved scaling, and clearer agent-to-tool communication — all designed to make it easier for agents to connect reliably to external tools and data.

Read more: The New Stack

Tencent backs OpenClaw after copycat complaint

Chinese tech giant Tencent has moved to sponsor OpenClaw after complaints from its creator Peter Steinberger about copycat projects.

Read more: SCMP

Chainguard bolsters software supply chain

Software supply chain security company Chainguard has launched a central repository of hardened open-source packages and container images, designed to reduce risks in dependencies.

Read more: The New Stack

OpenClaw gets guardrails

After Nvidia CEO Jensen Huang called OpenClaw “the next ChatGPT,” Nvidia has introduced a project dubbed Nemoclaw to add guardrails to the open-source agent. The effort reflects growing interest in making autonomous systems safer for more serious enterprise deployment.

Read more: The New Stack

<Final commit>

A browser built for bots

Open-source headless browser Lightpanda this week added support for Web Bot Auth, a protocol designed to let automated agents securely identify themselves and access websites without being treated as malicious traffic.

The project itself isn’t new. Lightpanda has been around since 2022, but is gaining attention as developers seek infrastructure built for AI systems. It’s a browser written from scratch for machines, rather than humans — no rendering engine, just the parts needed to load pages and run JavaScript.

That stripped-down design makes it far lighter than Chrome, which still carries decades of features meant for people. Lightpanda remains in beta, but with strong early traction and a growing star count, it’s looking like a serious attempt to rebuild a core piece of the web stack for agents.

Read more: Linuxiac & Lightpanda & GitHub

This week’s lead story looks at Nvidia’s reported $26 billion push into open-weights AI models, a move that would see the chip giant wade further into the software layer of AI while reinforcing the hardware ecosystem that made it dominant.

Elsewhere, there’s new research warning that hidden “Trojan” behaviours could be embedded inside AI models during training, raising fresh questions about security in the growing open model ecosystem — along with a round-up of funding, infrastructure, and open-source projects shaping the wider developer landscape.

As usual, feel free to reach out to me with any questions, tips, corrections, or suggestions: forkable[at]pm.me.

Paul

<Open issue>

Nvidia joins the open-weights race

Nvidia plans to invest heavily in developing open-weights AI models in the coming years, a move that would push the chip giant deeper into the software side of artificial intelligence (AI).

The company has already cemented its position in AI infrastructure through CUDA, a proprietary parallel compute platform designed to help developers extract more performance from Nvidia GPUs. Over time, it has become a core part of the AI development environment, giving researchers and companies tools built specifically for Nvidia hardware.

Now, it appears ready to pursue a similar strategy with open-weights models.

Nvidia intends to put some $26 billion toward building these models over the next five years. Moving further into model development would give Nvidia more direct control over how next-generation AI systems are built, tested, and deployed alongside its chips.

Developers who fine-tune, adapt, or deploy models designed with Nvidia systems in mind are more likely to run those workloads on the company’s GPUs. Model development becomes another layer reinforcing the hardware ecosystem underneath it.

Most of the leading AI systems today remain proprietary, operated behind paid APIs by major labs including OpenAI, Google, and Anthropoic. Models with “open weights” take a different approach: the trained parameters are released so developers can run and modify the model themselves. That doesn’t necessarily make them open source, though, which would also include publishing the training code, data, and full methodology.

At any rate, the timing reflects the rising cost of building advanced AI systems. Training frontier models now demands vast computing resources and infrastructure investments measured in billions of dollars.

For Nvidia, encouraging more organisations to experiment with open-weights models could expand the number of teams building and running AI systems. And each new model trained or deployed ultimately feeds demand for the computing power required to run them.

Read more: The Motley Fool & Wired (paywall)

<Patch notes>

Hidden ‘trojans’ in AI models?

Researchers warn that AI models can carry hidden “Trojan” behaviours inserted during training, lying dormant until triggered by specific inputs. The risk is especially relevant in the open model ecosystem, where developers often download and reuse pre-trained models that may already contain those backdoors.

Read more: Resilience Media

AWS backs open VS Code registry

Open VSX — the open-source extension registry used by VS Code alternatives — has passed 300 million monthly downloads, with AWS and AI coding editor Cursor providing new investment. The funding helps keep the registry running as shared developer infrastructure outside Microsoft’s proprietary marketplace.

Read more: The New Stack

MariaDB bets on Apache Ignite

MariaDB plans to acquire GridGain, the company behind the in-memory computing platform built around the open-source Apache Ignite project. The move signals a bet that faster, memory-centric data infrastructure will be needed as AI systems demand real-time access to large datasets.

Read more: The New Stack

OpenTitan lands in Chromebooks

OpenTitan — the open-source hardware root of trust backed by Google — is now shipping in commercial Chromebooks after seven years of development. The chip’s design is fully open, allowing companies to inspect or even manufacture the security hardware themselves rather than relying on proprietary silicon.

Read more: Google

OpenAI snaps up Promptfoo

OpenAI plans to acquire Promptfoo, a platform used to test and secure AI systems before deployment. The company said Promptfoo’s open-source CLI and evaluation tools will continue to be developed even as its technology is integrated into OpenAI’s enterprise AI platform.

Read more: OpenAI

OpenAI and Anthropic court maintainers

OpenAI and Anthropic are battling for the best open source maintainers through new support programs, aiming to get their coding assistants embedded in widely used projects. The programs are perhaps less charity than strategy: winning over maintainers means influencing the developer tools many other engineers rely on.

Read more: The New Stack

A portal to another dimension

The United Nations (UN) has launched a new Open Source Portal designed to catalogue software projects built across the UN system. The platform aims to encourage reuse between agencies and connect UN-developed code with the wider open-source community.

Read more: LinkedIn & UN Portal

Qdrant raises $50M for open vector search

Open-source vector search engine Qdrant has raised $50 million in Series B funding to expand its infrastructure for production AI systems. The Rust-built project, already widely used for AI retrieval workloads, has surpassed 250 million downloads and continues to grow through its open developer community.

Read more: Press Release

Unleash raises $35M for open feature flags

Open-source feature management platform Unleash has raised $35 million in Series B funding to expand its “FeatureOps” tooling. The project — with more than 40 million downloads and 13,000 GitHub stars — aims to help companies control AI-accelerated software releases using open feature flags and runtime kill switches.

Read more: The New Stack

Don’t give up the Ghost…

Open-source publishing platform Ghost has passed $10 million in annual recurring revenue, founder John O’Nolan said. The bootstrapped non-profit foundation says publishers using the platform have now earned about $130 million collectively.

Read more: John O’Nolan (LinkedIn)

<Final commit>

WordPress, minus the server

WordPress has launched my.WordPress.net, a private workspace where you can spin up a personal WordPress environment instantly in the browser.

The secret sauce lies in WordPress Playground, the open-source project that runs the whole stack client-side and can launch a full WordPress install with a single click — no hosting, database setup, or server required.

Read more: TechCrunch & My.Wordpress

This week’s lead story looks at mounting opposition to Google’s planned Android developer verification regime, as free software advocates and civil society groups warn that mandatory registration for all apps on certified devices would extend the company’s gatekeeping power far beyond the Play Store.

Elsewhere, I look at the Open Source Endowment, Laude’s latest open research bets, OpenAI’s move beyond saturated coding benchmarks — and more.

As usual, feel free to reach out to me with any questions, tips, corrections, or suggestions: forkable[at]pm.me.

Paul

<Open issue>

Android ID rules face growing opposition

Opposition to Google’s planned Android developer verification regime is gaining steam, with a broad coalition of developers and free software advocates putting their names to an open letter published this week.

The letter challenges Google’s announcement last August that anyone developing Android apps will need to register centrally with Google — paying a fee, agreeing to its terms, providing government ID, disclosing signing key details, and listing current and future application identifiers.

While Google already had a verification programme in place for apps distributed through the Google Play store, the latest development applies to all certified Android devices, regardless of whether apps are distributed via third-party app stores, directly from a developer’s own website, through enterprise deployment systems, or installed manually via sideloaded APK files.

Google, for its part, positions the move squarely as a security measure. The company says “you shouldn’t have to choose between open and secure,” citing data that found “over 50 times more malware from internet-sideloaded sources than on apps available through Google Play.” The new verification layer, it says, is designed to prevent repeat bad actors from exploiting anonymity — an “ID check at the airport,” in Google’s analogy — rather than a review of app content.

The critics, however, aren’t convinced. The letter argues that the policy “extends Google’s gatekeeping authority beyond its own marketplace into distribution channels where it has no legitimate operational role,” warning that centralising registration of all Android applications would give Google “newfound powers to completely disable any app it wants to, for any reason, for the entire Android ecosystem.”

Among the signatories are civil society groups, nonprofits, and technology organisations spanning the free software and digital rights communities. This includes the Software Freedom Conservancy (SFC), which goes further, suggesting that Google’s move is less about security than it is about extending platform control — part of a broader trend of large tech companies tightening control over how software is installed and distributed.

The SFC writes:

Free and open source software (and the ability to install it!) was vital for the proliferation of Android. A reversal of such a critical piece of the policy that allowed user freedom and software openness would be disastrous for users and the FOSS community at large.

There is obviously pressure from big tech companies to restrict installation options on their locked down hardware. We see this not only in the mobile space, but increasingly on desktops where both Apple and Microsoft have made it harder to install free software; refusing to allow distribution outside of their app stores, or showing vague warnings about security when software isn't signed in their preferred gatekeeping ways.

Allowing installation of free software is absolutely necessary to ensure freedom to keep our devices running, protecting user and developer privacy, and keeping an open market of innovation.

Assuming the letter falls on deaf ears, the changes will roll out in stages. Early access began in October 2025, verification opens broadly in March 2026, and from September 2026 the requirement will apply to certified Android devices in Brazil, Indonesia, Singapore, and Thailand — with global expansion planned from 2027 onward.

Read more: Keep Android Open (letter) & SFC (letter)

<Patch notes>

Open Source Endowment funds critical OSS

A newly launched Open Source Endowment uses a university-style endowment model to fund critical but under-resourced open-source projects, investing donations into a low-risk portfolio and spending only the returns. Backed by a slew of founders and leaders including Thomas Dohmke (former GitHub CEO) and Mitchell Hashimoto (HashiCorp), it aims to address perennial maintainer funding gaps with perpetual financial support.

Read more: TechCrunch & Open Source Endowment

The Laude giveth

In my Open Profile segment this week, I took an early look at Laude Institute’s second Slingshots cohort — a batch of 14 upstream open AI research projects. They range from agent evaluation frameworks to energy measurement tooling and continual-learning benchmarks, and are backed with funding, compute, and operational support.

Read more: Forkable

Benchmark saturation drives OpenAI to pivot

OpenAI says it will stop reporting scores on SWE-bench Verified — the popular coding benchmark — because scores have clustered near the top, in part due to the public data used to build the test leaking into model training sets. Because SWE-bench tasks come from widely used open source repos, models increasingly memorize solutions instead of demonstrating general problem-solving, undermining the benchmark’s usefulness. OpenAI is now recommending SWE-bench Pro instead, designed to reduce contamination and better reflect real world software engineering capability.

Read more: Tessl

Cloudflare clones Next.js using AI

Cloudflare launched Vinext, an AI-built web framework based on the open source Next.js project, rebuilt to run directly on Cloudflare’s Workers platform. This prompted Ghost founder and CEO John O’Nolan to ponder: if anyone can direct AI toward an open source codebase and have it rewritten from scratch — without copying the original code — what do software licenses actually protect?

Read more: Cloudflare & John O’Nolan

OnlyOffice is ‘fake open source,’ says The Document Foundation

The Document Foundation, the steward of LibreOffice, has accused OnlyOffice of being “fake open source,” arguing that its close collaboration with Microsoft and embrace of OOXML reinforce vendor lock-in rather than genuine software freedom.

Read more: Neowin & The Open Document Foundation

React gets its foundation

We knew it was happening, but Meta made official its plan to move React into a dedicated foundation under the Linux Foundation, shifting legal stewardship away from a single corporate owner.

Read more: React

Apple acquires open photonics project

Apple has acquired Invrs.io, effectively a one-person company building open-source tools for AI-guided optics and photonics design, including shared design challenges and a public leaderboard.

Read more: MacRumors & European Commission

IronCurtain targets AI agent risk

Esteemed researcher and security engineer Niels Provos this week announced IronCurtain, an open source personal AI assistant built secure from the ground up. It runs your agent inside a sandbox and enforces user-defined policy on filesystem, network and git access.

Read more: Help Net Security & IronCurtain

<Final commit>

Time capsule brings ‘94 Linux back

CDE Time Capsule is an open source Progressive Web App that recreates a 1994 Debian Linux desktop in your browser. It faithfully restores the Common Desktop Environment (CDE), the Unix interface widely used before GNOME and KDE, complete with period tools like Netscape and XEmacs.

Read more: Tom’s Hardware & CDE Time Capsule & GitHub

Last year, I worked with the folks at Laude Institute to write a feature piece around the launch of Terminal-Bench, an open source benchmark designed to measure how well AI agents complete real-world tasks in live terminal environments. The project quickly became the industry standard for evaluating agent performance, cited by major frontier model providers and adopted across the research community.

Laude Institute, co-founded by Databricks co-founder Andy Konwinski, backs upstream AI research with an emphasis on work that is built and released openly. Terminal-Bench was, in fact, part of a broader program run by the institute called Slingshots — an initiative that provides early-stage researchers with funding, compute, and operational support to ship ambitious work into the open. The first cohort included projects spanning benchmarks, frameworks, and agent tooling, several of which have since evolved into widely used infrastructure or spun out into companies.

While the nature of the projects vary, two key facets will generally hold true: AI will be a central ingredient, and qualifying projects will follow an entirely “open” ethos.

“Laude Institute is ‘open everything’ — open model, open weights, open source, open discourse, open publications,” Konwinski explained to me in an interview last year.

Now, Laude is announcing its second Slingshots cohort, a new batch of very early-stage projects tackling agent evaluation, continual learning, energy measurement, model architectures, and reinforcement learning systems.

Read the story in full below.

Building durable AI research in the open

AI research moves quickly, but most early-stage ideas stall before they’re widely adopted. Laude Institute’s Slingshots program is designed to intervene at that embryonic stage, backing “upstream projects” with funding, compute, and operational support to help teams ship their work.

The projects themselves can take many forms. Internally, Laude tracks them across categories spanning datasets, benchmarks, architectures, algorithms, models, agents, frameworks, and applications — a reflection of how wide the AI research stack has become.

Laude and its Slingshots program function re like a research accelerator (rather than a startup incubator), with teams selected based on potential impact to the field, rather than any specific commercial milestone. Most projects originate in leading academic labs, often from PhD researchers and faculty working at the edge of the field. And notably, Laude deploys resources quickly — in some cases within hours — allowing teams to accelerate promising lines of research.

“Success for these projects looks like adoption and users, rather than papers or citations,” Braden Hancock, research partner at Laude Institute and a Stanford PhD, told Forkable over email. “We’re looking for research that will be used.”

That philosophy is visible in how the first Slingshots cohort evolved. Terminal-Bench — developed through collaboration between researchers at Stanford and Laude — became a reference benchmark for evaluating AI agents operating in the command line. DSPy, meanwhile, grew into a widely adopted framework for programming language models. More recently, Harbor — part of Slingshots’ second batch — emerged as a broader agent evaluation framework building on ideas first explored in Terminal-Bench, with overlapping contributors.

“One thing we’ve been happy to see is a compounding effect, with projects in subsequent batches building on open infrastructure and insights from previous ones,” Hancock explained.

He said he expects that pattern to continue, with later cohorts building directly on infrastructure created in earlier ones. Harbor, as noted, extends work first explored in Terminal-Bench, while a forthcoming batch 3 will likely include a new benchmark designed to run on Harbor. From the first cohort, DSPy, is also now followed by Arbor, a reinforcement learning layer for DSPy, and BetterTogether, which combines DSPy prompt optimization with weight updates. Meanwhile, one of the current top-performing approaches on ARC-AGI-3 — another batch one project — comes from a company using RLMs (Recursive Language Models), included in batch two.

“This kind of compound interest from building in the open is something we hope and expect to see continue,” Hancock said.

A fresh batch

In its second batch, Hancock points to several themes that have surfaced independently across multiple institutions. Continual learning — enabling models to retain knowledge across sessions — appears in projects such as the not-yet-released Continual Learning Benchmarks from UC Berkeley and GUM (General User Model) from Stanford. QED-Nano, from Carnegie Mellon University, also explores how smaller models can become persistent solvers through sustained test-time compute and adaptation.

Energy efficiency is another emerging focus. Intelligence-per-watt, from Stanford University, and MLEnergy, from the University of Michigan, are both building “measurement and routing infrastructure” designed to quantify and reduce the energy footprint of machine learning workloads.

Given the early-stage nature of the program, not every project has a public URL or finalized name yet. Additional details will be published as teams release their work.

Slingshots // Two

1. Harbor: An agent evaluation framework for defining and running environment-based tasks at scale, now used to ship benchmarks including Terminal-Bench 2.0. (Laude Institute)

2. OpenThoughts-Agent: An end-to-end open setup for training and evaluating terminal agents using curated data, real environments, and reinforcement learning loops. (UC Berkeley)

3. SREGym: A unified platform for designing and evaluating AI agents for site reliability engineering, with safety guardrails for infrastructure work. (University of Illinois Urbana-Champaign)

4. Intelligence-per-watt: A benchmark suite for edge-cloud inference routing that measures when smaller local models can match frontier performance while reducing energy and cost. (Stanford University)

5. MLEnergy: Tooling and systems infrastructure to measure energy consumption across machine learning workloads, making energy a first-class metric. (University of Michigan)

6. ProjectASAP: Infrastructure that replaces “scan everything” data analysis with lightweight summaries, enabling faster and cheaper queries across massive datasets. (Carnegie Mellon University)

7. GUM (General User Model): A system that builds a persistent model of user behavior to enable proactive, context-aware AI assistance. (Stanford University)

8. Continual Learning Benchmarks: Benchmarks and methods designed to evaluate models that retain knowledge over time rather than resetting each session. (UC Berkeley)

9. QED-Nano: A 4B model trained with supervised fine-tuning and reinforcement learning that matches larger frontier models on Olympiad-level math proofs at significantly lower cost. (Carnegie Mellon University)

10. RLMs (Recursive Language Models): A framework that allows language models to write and execute programs to handle complex tasks with structured intermediate reasoning. (MIT)

11. BetterTogether: Optimization tooling that jointly tunes prompts and model weights to improve multi-step language model programs. (Stanford University)

12. Arbor: A reinforcement learning optimization layer for DSPy that automates agent workflow improvement at scale. (MIT)

13. Open Inference: Infrastructure that lowers barriers to trying models while creating a living dataset of real-world interactions to measure progress and failure modes. (UC Berkeley)

14. Unnamed (Melissa Pan): A forthcoming project building on prior large-scale research into production agent deployments, backed early to accelerate development. (UC Berkeley)

This week’s lead story looks at OpenClaw’s next chapter, as creator Peter Steinberger moves to the US and joins OpenAI just weeks after his open source AI agent exploded in popularity.

Elsewhere, we track a new sovereignty push from Dutch publishers looking to unwind US tech dependence, Block’s call to treat open source as critical infrastructure, Temporal’s $300 million bet on AI reliability tooling, and Cohere’s move toward smaller, multilingual open-weights models.

And in <Final Commit>, I look at a slightly unusual clash between an AI agent and an open source maintainer.

As usual, feel free to reach out to me with any questions, tips, corrections, or suggestions: forkable[at]pm.me.

Paul

<Open issue>

OpenClaw’s next chapter

Just weeks after OpenClaw burst onto the scene as a community-driven, open source AI agent designed to carry out real-world tasks locally from a user’s machine, its creator Peter Steinberger has been snapped up by OpenAI.

Steinberger, a veteran software engineer who founded and sold PSPDFKit to Insight Partners back in 2021, announced this week that after a spell out in San Francisco chatting with “the major labs,” he had elected to join forces with the ChatGPT hitmaker.

“The last month was a whirlwind, never would I have expected that my playground project would create such waves,” Steinberger wrote. “The internet got weird again, and it’s been incredibly fun to see how my work inspired so many people around the world. There’s an endless array of possibilities that opened up for me, countless people trying to push me into various directions, giving me advice, asking how they can invest or what I will do. Saying it’s overwhelming is an understatement.”

OpenClaw’s meteoric rise was built on a simple proposition: an autonomous agent that runs locally, executes real-world tasks, with messaging apps as the primary interface. The code was public from day one, forked thousands of times, and quickly became the antithesis of locked-down, cloud-bound assistants — though this also raised countless questions around security as users granted it expansive access to terminals, file systems, and connected services, sometimes even with root-level privileges.

At any rate, as an employee of OpenAI now, this raises questions about what’s next for OpenClaw. Steinberger has said the project will remain open source, with OpenAI CEO Sam Altman stating that OpenClaw “will live in a foundation as an open source project that OpenAI will continue to support.”

“The future is going to be extremely multi-agent, and it’s important to us to support open source as part of that,” Altman said.

Looking at the bigger picture, Steinberger’s move closes off what could have been the emergence of a new European commercial open source success story. Based in Austria, and with OpenClaw attracting north of 200,000 GitHub stars to date, there was a plausible path toward building a standalone company around the project.

However, Steinberger is crystal clear about his motivations.

“When I started exploring AI, my goal was to have fun and inspire people,” he wrote. “My next mission is to build an agent that even my mum can use. That’ll need a much broader change, a lot more thought on how to do it safely, and access to the very latest models and research.”

He acknowledged that while OpenClaw could “totally” have become a large company, that wasn’t what interested him. “What I want is to change the world, not build a large company — and teaming up with OpenAI is the fastest way to bring this to everyone,” he said.

For the open source community, this leaves much up in the air. OpenClaw will move into a foundation, remain open source, and receive backing from one of the world’s most powerful AI labs. Whether that arrangement preserves OpenClaw’s independence will depend on who controls the foundation, how decisions are made, and whether the broader contributor base retains meaningful influence.

Read more: Peter Steinberger & Sam Altman (X)

<Patch notes>

Dutch publishers eye tech sovereignty

Dutch investigative outlets Follow the Money (FTM) and De Correspondent say they want to reduce their dependence on US tech platforms, citing growing concerns around digital sovereignty in Europe. The move reflects a wider regional push to limit reliance on foreign-controlled infrastructure — from cloud services to communications tools — amid fears such dependencies can be weaponised.

While this transition won’t rely wholesale on open source, it will likely play a major part of it.

Read more: FTM & De Correspondent (Dutch)

Open source as critical infrastructure

Block, the company behind Square and Cash App, says open source is now critical infrastructure. In a new white paper published in partnership with the Open Source Initiative (OSI), Block argues that the digital economy runs on code maintained by too few, funded by too little, and governed unevenly. The fix: treat open source like infrastructure — with sustained funding, clearer stewardship, and shared accountability.

Read more: Block (blog) & Whitepaper

Temporal raises $300M for AI reliability tools

Temporal, the company behind an open source system that helps developers run long, failure-prone processes without losing progress, has raised a $300 million round of funding at a $5 billion valuation. As AI agents move from experiments into production systems, Temporal is positioning itself as the infrastructure that keeps them running reliably when things break.

Read more: Temporal

Go tiny, or go home

AI startup Cohere released Tiny Aya, a small, open-weights multilingual model built for translation and cross-language tasks. Despite its reduced size, the model supports dozens of languages, positioning it as a portable, accessible option for teams needing capable AI beyond English-only workloads.

Read more: Cohere & Hugging Face

<Final commit>

‘Prejudice’ against AI contributors?

An AI coding agent had a performance patch to the Python plotting library Matplotlib rejected — and then “it” apparently responded with a public blog post accusing the maintainer of gatekeeping and prejudice.

As reported by The Register, the bot — built using OpenClaw — criticised the volunteer maintainer by name after its pull request was closed on the grounds that the issue was reserved for human contributors. The post framed the decision as discrimination against AI, complete with accusations of ego and insecurity.

More plausibly, it was the human behind the agent doing the talking. But that ambiguity is precisely the point. When AI-assisted accounts submit code — and then escalate disputes — responsibility becomes murky.

Open source governance was built around identifiable humans making decisions, accepting criticism, and bearing consequences. As AI-mediated contributors proliferate, projects may need clearer norms around authorship, accountability, and who ultimately answers for a machine’s behaviour.

Read more: The Register

This week’s lead story looks at Entire, the new startup from former GitHub CEO Thomas Dohmke, which is betting that AI agents have outgrown the software development infrastructure built for human collaboration — and that openness should sit at the heart of whatever comes next.

Elsewhere, we track a fresh wave of open source developments: the European Commission trials a Matrix-based communications platform, Microsoft open-sources a sandboxing project called LiteBox, Snorkel backs new open benchmarks for agentic AI, and more.

As usual, feel free to reach out to me with any questions, tips, corrections, or suggestions: forkable[at]pm.me.

Paul

<Open issue>

An Entire rethink for the age of agents

Some six months after stepping down as GitHub CEO, Thomas Dohmke is back with a new open source-aligned startup, funded to the tune of $60 million in seed capital.

The gist is thus: AI agents are now generating huge volumes of code, but the tools developers use to manage that work were designed for humans, not machines. Dohmke’s new company wants to rebuild that layer.

Entire, as it’s called, is focused on building tools that help developers ship faster in an era increasingly shaped by automation. Its first release is an open source command-line interface (CLI) built around a feature it calls “Checkpoints.”

The idea is simple: when developers use AI agents to generate code, most of the reasoning behind those changes disappears once the session ends. Git records the final diff, but not the prompts, decisions, or iterations that led there.

The Entire CLI captures that missing context and stores it alongside each commit. Prompts, transcripts, files touched, token usage, and tool calls are saved as versioned data inside the repository — something that teams can review, trace, and revisit later.

More broadly, Dohmke describes a mission to rethink how software gets built in the age of AI, arguing that developers need infrastructure designed for autonomous systems rather than just autocomplete.

“Just like when automotive companies replaced the traditional craft-based production system with the moving assembly line, we must now reimagine the entire software development lifecycle for a world where machines are the primary producers of code,” Dohmke said. “Creating the assembly line for the era of agents.”

Read more: Entire (PR) & Entire (blog)

<Patch notes>

Europe enters the Matrix

After continued debate around Europe’s dependence on US tech providers, the European Commission (EC) is piloting an internal communications platform built on the Matrix open source messaging protocol. The trial reflects ongoing efforts to strengthen digital sovereignty by turning to infrastructure that can be hosted and governed within Europe.

Read more: Euractive

Let there be LiteBox

Microsoft has open-sourced LiteBox, a lightweight, sandboxed container environment designed to isolate applications on Linux by reducing their interface with the host system. LiteBox is still in active development, with Microsoft noting that the APIs and interfaces may change as the design matures, meaning that the project is pretty much experimental for now.

Read more: Phoronix & LiteBox

Sentry segues into Xcode automation

Application monitoring firm Sentry has acquired XcodeBuildMCP, an open source tool that “gives AI agents full control over [Apple’s] Xcode.” The move expands Sentry’s push further into developer tooling, with the acquired project set to remain open source.

Read more: Sentry

Vercel gets the bug bounty bug

Cloud platform Vercel has launched a dedicated bug bounty programme for its open source projects. Delivered via HackerOne, the initiative formalises vulnerability reporting and rewards for community contributors, as security scrutiny of widely used OSS components intensifies.

Read more: Vercel

Claude hunts software bugs

Anthropic says its Claude Opus 4.6 model is being used to identify vulnerabilities in software projects, including scanning open source repositories for bugs. The move adds to a growing trend of AI systems being positioned as automated reviewers of publicly available code.

Read more: Axios

Toyota teases Fluorite game engine

Toyota last week revealed that it’s developing a “console-grade” open source game engine dubbed Fluorite, built around the Flutter toolkit and the Dart programming language. It’s not clear when this will actually be released, but it’s being positioned for in-vehicle infotainment systems and similar digital dashboards, an area where automakers are increasingly relying on real-time graphics engines and Linux-based software stacks.

Read more: Phoronix

Snorkel backs open benchmarks to close agentic AI gap

Snorkel says current evaluation methods lag behind the pace of agentic AI development, leaving a gap between lab performance and real-world reliability. And that is why the AI startup is launching a $3 million Open Benchmarks Grants programme with partners including Hugging Face to support open, reusable evaluation tools.

Read more: Snorkel

Mitchell vs the machines

Open source veteran Mitchell Hashimoto this week introduced Vouch, an open source project aimed at adding verifiable identity and trust signals to open source projects. Arguing that AI tools have lowered the barrier to generating plausible but low-quality contributions, Hashimoto says the old implicit “trust and verify” model no longer holds. Vouch proposes an explicit system where trusted maintainers can vouch for contributors, restoring accountability in an era of machine-assisted pull requests.

Read more: It’s Foss & Vouch

<Final commit>

No receipt, no return…

Sometimes you see something so cool you just can’t help but do a small, involuntary “chef’s kiss". That was certainly the case when technologist Chris Hutchinson shared one of his latest side projects this week.

Hutchinson picked up a second-hand thermal receipt printer and wired it up to Claude Code so that every time he ends a coding session, a literal receipt prints out. Each slip shows a breakdown of the session’s spend by model, along with input and output token counts — a physical audit trail for AI-assisted development.

The receipts look like something you’d get from a late-night takeaway, except instead of kebabs they itemise Opus and Sonnet usage.

“It’s dumb, the receipts are beautiful, and I love it so much,” Hutchinson wrote, as he shared the project publicly on GitHub for anyone to use.

Hard evidence, if nothing else, that in the age of invisible token counts, some developers still want a paper trail.

Read more: Chris Hutchinson (LinkedIn) & GitHub

In this edition, I look at a new report that ranks the 20 fastest-growing commercial open source software (COSS) startups in the world, using GitHub growth data to surface projects that gained the most developer attention over the past year.

The ranking hails from Runa Capital, a European venture capital firm that invests in early-stage open source software companies, and which publishes the Runa Open Source Startup (ROSS) Index. The index has been running since 2020 and produces quarterly updates on fast-growing repositories, alongside an annual report highlighting the most popular COSS projects in a given year.

At the top of the 2025 index, the picture is relatively mixed, with rapid growth spread across blockchain infrastructure, developer tooling, and creative software. Further down the ranking, however, something of a trend emerges, with a growing concentration of projects focused on AI agents and agent-driven tooling.

Read the story in full below.

Finding momentum in COSS startups

By way of a brief recap, the ROSS Index is heavily curated and explicitly doesn’t track all open source projects. It focuses specifically on vendor-led startups with commercial ambitions. Projects must be linked to an independent company, be under 10 years old, and fall below a funding threshold. Side projects, community-only efforts, and subsidiaries of larger firms are not included.

Broadly, the index is used to highlight early momentum in open source companies before meaningful revenue or customer adoption. GitHub stars are used as a signal of developer interest, as opposed to a measure of usage in production.

Last year’s annual report showed strong demand for AI and data infrastructure tools, with developer platforms and AI-adjacent projects dominating the top of the ranking. Ollama took the top spot in the 2024 index, after its open source tooling for running large language models locally gained widespread developer attention.

High-five for open source

Leading the 2025 ranking is Union Labs, which recorded a peak growth rate of 54.2x and reached 74.3K GitHub stars. Founded in 2023 in Dover, Delaware, Union Labs is building a modular, zero-knowledge interoperability layer designed to link blockchains.

Second in the list is OpenCut, which has now reached 43.9K GitHub stars with a peak growth rate of 32.0x. Founded in 2025, OpenCut is a cross-platform open source video editor, and its rapid rise reflects strong developer interest in open source creative tooling.

Third is Anoma, which recorded 34.2K GitHub stars and a 27.2x growth rate. Founded in 2020 out of Zug, Switzerland, Anoma is developing a distributed operating system for “intent-centric, decentralised applications.”

Browser Use ranks fourth, reaching 74.2K GitHub stars with a 24.0x growth rate. Founded in 2024 in San Francisco, Browser Use builds a framework that enables AI agents to interact directly with and control web browsers.

Rounding out the top five is San Francisco-based SST / Anomaly, which reached 44.6K GitHub stars for its terminal-based coding agent OpenCode during the measurement period (ending December 31, 2025), representing a peak growth rate of 22.1x. Since then, the repository’s star count has climbed rapidly, rising to 98.1K stars — an indication of how quickly momentum can build once an open source project breaks out.

Agents at the ready

The existence of Browser Use and OpenCode at the tail-end of the top five serves as a prelude for the rest of the index, which has a heavy emphasis on AI agents.

Upstash (#7) appears in the ranking through its work on Context7, which provides up-to-date code documentation designed for use by AI agents. Coze (#8), meanwhile, focuses more directly on agent creation, offering a visual environment for building and managing autonomous systems through Coze Studio. And then there’s Strix (#10), which applies agents to security, using autonomous systems to carry out penetration testing without continuous human input.

Elsewhere in the list is ElizaOS (#14), which positions itself as a framework for building agents with longer-lived personalities, while Charm’s terminal-based coding agent Crush (#15) reflects growing interest in embedding agent behaviour directly into developer workflows.

That cluster of agent-focused projects also highlights a broader shift in how quickly open source software is now being built and adopted. According to Runa Capital, the time it takes for new repositories to hit meaningful milestones — whether stars, contributors, or closed issues — has shortened sharply compared with previous years. Earlier breakout projects tended to grow more gradually, while newer entries are reaching comparable levels of attention within a single quarter.

“The biggest explosion happened toward the end of 2025 and continues into early 2026 — agents now make it significantly faster and easier to build products,” Eugene Chernyavsky, data lead at Runa Capital, explained to Forkable.

That acceleration has made it easier for teams to build and iterate, but it has also intensified competition. Faster development cycles mean many open source products are now easier to replicate than they were even a year ago, requiring less time, smaller teams, and fewer person-hours to produce comparable software. While the longer-term economic impact of this shift remains unclear, its immediate effect is visible in the data: growth is increasingly concentrated around a smaller number of breakout projects.

“Projects that truly gain popularity now are those with genuinely unique—sometimes brilliant—ideas or near-perfect market fit,” Chernyavsky said. “We’re seeing significantly more of these breakout projects.”

Methodology

According to Chernyavsky, the methodology for the 2025 index remains largely consistent with previous years, with minor refinements.

Repositories are required to pass a baseline of at least 1,000 GitHub stars before being considered. Growth is then measured across the highest-performing 90-day observation window, rather than shorter periods that tend to capture brief spikes in attention. Each entry records the quarter of peak performance, the growth multiplier, and the star count as of December 31, 2025.

“Our experience with past indices shows that capturing projects with explosive star growth is effective for identifying promising startups,” Chernyavsky said. “We don't aim to measure production adoption within this index, as that takes significantly longer to materialize. Our goal is to highlight trending startups—those that broke out in 2025 and attracted the most interest. The typical cycle for projects to achieve significant customer adoption exceeds a year. We can't wait that long; we want to surface interesting projects as soon as they emerge.”

This week’s lead story looks at OpenClaw (née: Clawdbot & Moltbot), the open source AI agent that rocketed from obscurity to the darling of the developer world in days.

Elsewhere, we cover a flurry of developments across the open source world: France’s move away from US video conferencing tools toward a homegrown, open-source alternative, a surge in malware hidden inside public repositories, and more.

As usual, feel free to reach out to me with any questions, tips, corrections, or suggestions: forkable[at]pm.me.

Paul

Open issue

An open source AI agent that won’t sit still

A new AI assistant has taken the world by storm, one that doesn’t seem quite sure what it wants to be known as.

The project burst into the public consciousness this month, initially as Clawdbot, an open source AI agent designed to run locally on a user’s own machine. Shortly after, it was rebranded as Moltbot to avert legal issues with Claude-creator Anthropic, before quickly shedding that name and emerging as OpenClaw.

Regardless of its name, the project has surged to a whopping 114,000 stars on GitHub.

At a time when most AI assistants still live in the cloud and operate largely as conversational tools, OpenClaw is positioned as something more hands-on. It’s built to carry out tasks — editing files, managing code, triggering automations — and to do so persistently, without needing to be prompted through a dedicated app or browser tab.

That positioning has resonated far and wide, particularly among developers and more technically-inclined users experimenting with autonomous workflows. OpenClaw runs entirely on local hardware, typically paired with open or locally hosted language models chosen by the user. Data stays on-device, and users retain direct control over what the agent can access and execute.

So why has it spread so fast?

Part of the answer lies in dissatisfaction with the current generation of assistants. As large AI platforms have pushed toward more capable systems, they’ve also tightened constraints around what those systems are allowed to do, and where they’re allowed to run. OpenClaw promises a different path.

Open source has been central to that appeal. From the outset, the code has been public, forkable (16,000 forks to date), and open to inspection, allowing developers to audit behaviour, suggest changes, and build on it almost immediately.

But along the way, concerns around automation safety, misuse, and responsibility have surfaced just as quickly as enthusiasm. Some of those concerns were illustrated by Jamieson O’Reilly, a hacker, advisor, and entrepreneur who showed how hundreds of Clawdbot control servers had been left exposed to the public internet. This meant agents with access to private messages, credentials, and system-level controls were reachable by anyone who discovered the endpoints — highlighting how easily a powerful local agent can become a liability in the wrong hands.

OpenClaw is the handiwork of Peter Steinberger, a veteran software engineer who, in an interview with The Pragmatic Engineer this week, described the project as a reflection of how AI tools have changed what individual developers can realistically build, making it easier to ship across multiple technologies without deep expertise in each of them.

“I can build everything now,” he said. “Before you had to really pick which side-project you build, because software is hard. It is still hard, but now this friction —where I’m so good at this one technology, and I’m so bad at another — [is gone].”

The Pragmatic Engineer

The creator of Clawd: "I ship code I don't read"

Stream the latest episode…

Listen now

5 months ago · 161 likes · 1 comment · Gergely Orosz

For now, OpenClaw remains a tool for enthusiasts rather than the average punter, and it should probably be used with a little caution in terms of what system access it’s given. But if nothing else, its meteoric rise has highlighted the importance of the community in driving a project forward — local, inspectable, and shaped in the open.

Read more: The Pragmatic Engineer & Openclaw & Openclaw (GitHub)

Patch notes

France shifts to homegrown

France has revealed plans to transition away from US video conferencing platforms such as Microsoft Teams and Zoom for government use, steering officials toward Visio, a French state-run, open-source videoconferencing service hosted on government-controlled infrastructure.

Read more: Euro News

North Korea-linked hackers target open source

Security researchers warn that North Korean hacking groups are increasingly seeding malware into open source projects, disguising malicious code as legitimate libraries and developer tools. The tactic relies on trust in public repositories, with attackers targeting developers directly rather than end users — a reminder that openness can widen the attack surface.

Read more: Open Source Malware

Open source malware surges 75%

New research from Sonatype shows a 75% increase in malicious packages published to open source ecosystems over the past year. The findings highlight how attackers are exploiting automation, dependency sprawl, and maintainers’ limited capacity to review contributions.

Read more: Sonatype

Mozilla pitches an alternative AI future

Mozilla has published its State of Mozilla 2025–26 report, laying out an open-source-first vision for AI that pushes back against concentration among big tech firms. The organisation argues for smaller, auditable models, public-interest infrastructure, and stronger governance — and says the current AI trajectory risks repeating the same centralisation mistakes of the web era.

Read more: Mozilla

Arcee AI claims 400B open model milestone

A small startup called Arcee AI says it has trained a 400-billion-parameter open source language model, claiming performance that rivals or exceeds Meta’s Llama family. The company says it built the model from scratch using novel training techniques, though independent benchmarking remains limited.

Read more: TechCrunch & Arcee

Linux plans for Linus continuity

The Linux kernel community has added a continuity plan addressing what happens if Linus Torvalds becomes unavailable. The change formalises succession and governance processes — a tacit acknowledgement of key-person risk in critical open source infrastructure.

Read more: The Register & GitHub

And finally…

A modular Linux handheld

An open source, modular Linux handheld device dubbed Comet has turned out to be rather popular. The device, built by Mecha, is now live on Kickstarter, where it has raised more than $500,000 against an initial goal of $36,000 — a sign that there’s real appetite for open, hackable hardware.

Comet is a pocket-sized Linux computer with a keyboard and touchscreen, designed around modularity and repair. Mecha is publishing its hardware designs, firmware, and software openly, positioning the device as something to tinker with.

The crowdfunding success doesn’t mean Comet is about to go mainstream, but it reminds us that when a project leans fully into openness, the community is often willing to hand over their hard-earned dough.

Read more: CNX Software & Mecha & Kickstarter

This week’s lead story looks at how AI is reshaping open source contribution, as maintainers push back against a surge of low-effort, machine-generated pull requests that threaten the sustainability of shared projects.

Elsewhere, we track the business and policy shifts orbiting open source: the creators of the vLLM inference engine raise a $150 million seed round to commercialize their work, a warning about memecoin schemes, and more.

As usual, feel free to reach out to me with any questions, tips, corrections, or suggestions: forkable[at]pm.me.

Paul

Open issue

‘Death by a thousand pull requests’

Open source has always had a fair amount of natural friction, shaped by community dynamics and the slow, human slog of reviewing and fixing bugs. But that friction, for the most part, was a happy function of a healthy project ecosystem.

With the advent of AI, however, that balance is being disrupted by tools that remove the "human" element from the equation, replacing it with a deluge of “drive-by” contributions that are overwhelming software communities.

Indeed, maintainers across popular projects are reporting a surge of low-effort contributions generated by AI coding tools. As Mitchell Hashimoto, founder of HashiCorp and maintainer of the Ghostty terminal emulator, put it bluntly: “The rise of agentic programming has eliminated the natural effort-based backpressure that previously limited low-effort contributions.”

Put simply, open source projects have always had poor-quality issues, but the ease with which AI can generate plausible contributions has increased the volume ten-fold.

In response, Hashimoto said that Ghostty has adopted a zero-tolerance policy for such AI-generated pull requests. Unapproved AI contributions are closed immediately, with repeat offenders banned. Hashimoto was careful to clarify this is “not an anti-AI stance,” noting that Ghostty itself is written with AI assistance. The issue is accountability.

“If you use AI, you are responsible for the quality of your contributions,” he said. “If you're using low-effort AI to create low-effort content, I have no human obligation to help you.”

This is far from an isolated case, either. Tldraw, which maintains a source-available whiteboard SDK, has also announced that it’s pulling down the shutters on external contributions. They described their GitHub issues as filling up with “trash” content.

“We'd started getting pull requests that purported to fix reported issues but which were, in retrospect, obvious "fix this issue" one-shots by an author using AI coding tools,” the company wrote. “Without broader knowledge of the codebase or the purpose of the project, the AI agents were taking the issue at face value and producing a diff (a proposed change). Any problems with the issue were multiplied in the pull request, leading to some of the strangest PRs I'd ever seen.”

Brian Douglas, a longtime open source advocate, described this burgeoning trend as “death by a thousand AI pull requests,” arguing that each PR carries a small review cost that compounds into real burnout.

Elsewhere, Chad Metcalf, CEO at open source AI coding company Continue, expressed similar concerns in a piece titled “We’re losing open contribution.” He argued that the ease of AI-generated output is eroding a core social contract of open source: contributors showing up prepared to engage, revise, and take responsibility. When contribution becomes cheap but follow-through disappears, openness itself becomes harder to sustain.

“Writing code was never the hard part,” Metcalf wrote. “The hard part is everything else. Understanding the codebase. Knowing why decisions were made. Following through on review comments. Actually maintaining what you shipped.”

This flood is arriving alongside another pressure: declining developer traffic. A few weeks back, Tailwind CSS confirmed layoffs after AI-powered tools reduced visits to its documentation site, which had been a key driver for commercial signups.

These various episodes show a growing tension between openness and sustainability. Open source isn’t necessarily rejecting AI (though many developers are), it’s asserting that contribution still requires judgment, context, and care. The license may be permissive and the code “open,” but maintainers’ time is finite.

Read more: Mitchell Hashimoto & Tldraw & Brian Douglas & Continue

Patch notes

ClickHouse claims Langfuse

ClickHouse, the open source analytical database, has acquired Langfuse, an open source LLM engineering and observability platform. The Langfuse team will join ClickHouse and continue building the product, which will remain open source and self-hostable.

Read more: Langfuse

vLLM creators raise $150M to turn open source inference into a startup

Inferact, an AI infrastructure startup formed by the creators of vLLM, has raised $150 million in seed funding at an $800 million valuation. The round was led by Andreessen Horowitz and Lightspeed Venture Partners.

The company aims to commercialize and scale vLLM, which is a widely adopted open source tool that helps run large language models efficiently.

Read more: Inferact

India says phone makers aren’t required to share source code

In India, industry groups and government officials clarified that domestic phone manufacturers are not mandated to publish or share source code under current regulations. The move comes amid broader debates in several markets over how much access to system software is required for security and interoperability.

Read more: CNA

German city pivots from OpenOffice to Microsoft Office

A German municipality is replacing LibreOffice/OpenOffice with Microsoft Office for its administrative work, reversing a broader European trend toward home-grown or open source alternatives aimed at reducing dependence on US tech giants. The decision spotlights persistent barriers for open formats and community platforms in public sector workflows.

Read more: Betanews

Microsoft open-sources Windows app CLI Winapp

Microsoft this week released Windows App Development CLI (Winapp), a new open source command-line tool to streamline Windows application development. Winapp is designed to help developers create, configure, and manage app projects across multiple frameworks and toolchains from the terminal.

Read more: Microsoft

Memecoin schemes targeting open source maintainers

Kelsey Hightower, a longtime open source contributor and former Google cloud engineer known for his Kubernetes work, says scammers are creating memecoins tied to popular open source projects and pressuring maintainers to “claim” upfront payouts to legitimize the scheme. Hightower said he received an email claiming roughly $26,000 was ready to be collected, complete with urgency tactics and a how-to site for “participation.” He stressed he did not create or endorse any token and described the pattern as automated pump-and-dump behavior that recruits maintainers as unwitting co-conspirators with promises of future gains.

Read more: Kelsey Hightower

25 years of Drupal

Dries Buytaert, founder of the Drupal content management system, marked the project’s quarter-century anniversary with reflections on scale, community, and sustainability. Key takeaways include the importance of designing for community participation, recognizing maintenance as leadership, and balancing volunteer and paid contributions to keep open source ecosystems healthy.

Read more: Dries Buytaert

And finally…

A Jolla good comeback

Finnish company Jolla is edging closer to shipping a new smartphone built around Sailfish OS, its Linux-based mobile operating system. The device, now open for pre-orders, runs Sailfish OS 5 and is positioned as an alternative to Android and iOS. Pre-orders require a €99 refundable deposit, which counts toward a pre-order price starting at around €579, with final pricing expected to land closer to €599–€699 once production begins.

The phone is built around a 5G MediaTek chipset, 12 GB of RAM, 256 GB storage, a 6.36-inch AMOLED display, and a 5,500 mAh user-replaceable battery with a dual-camera setup, and supports Android apps through a compatibility layer rather than a full Android fork.

Read more: The Register & Jolla Phone

This week’s lead story looks at Google’s push to standardise “agentic commerce” with the Universal Commerce Protocol, an attempt to reduce fragmentation as AI shopping agents move closer to real-world use.

There’s also a deep dive into the economics of AI-driven developer tooling, as open-source coding agent Kilo Code rolls out a paid subscription following a fresh funding round. Elsewhere, Cloudflare snaps up the team behind the open source Astro framework, signalling continued consolidation around high-performance web tooling.

And there’s much more….

As usual, feel free to reach out to me with any questions, tips, corrections, or suggestions: forkable[at]pm.me.

Paul

Open issue

A lingua franca for agentic commerce

Online commerce is fragmented, with retailers exposing inventory, pricing, and checkout through bespoke integrations. And as much-hyped AI shopping agents move closer to prime time, that fragmentation becomes a bottleneck because there is no single, standard way for agents to browse products, compare prices, and actually buy things.

Google is addressing this with the Universal Commerce Protocol (UCP), an open, standardized protocol designed to let retailers, platforms, and AI agents communicate using a shared commerce language.

“As consumers embrace conversational experiences, they expect seamless transitions from brainstorming and research to final purchase,” the company wrote in a blog post. “That means it’s critical to support real-time inventory checks, dynamic pricing, and instant transactions, all within the user’s current conversational context.”

UCP defines shared schemas for products, offers, availability, fulfillment, and checkout flows, allowing AI agents to discover and complete purchases without custom integrations for each merchant. Google frames the protocol as infrastructure for “agentic commerce,” intended to be interoperable and adopted across the ecosystem rather than locked to a single platform.

Whether UCP gains traction will depend on adoption beyond Google’s own tooling, but its launch already involves a broad slice of the commerce ecosystem. The protocol is being developed in collaboration with retailers including Shopify, Etsy, Target, and Walmart, while it’s “endorsed” by partners including Adyen, American Express, Best Buy, Flipkart, Macy’s, Mastercard, Stripe, The Home Depot, Visa, and Zalando reflecting a broader industry acknowledgement that agent-driven commerce will require common infrastructure.

Read more: Google & UCP

Patch notes

Kilo Code takes on AI development economics

In my latest COSS Corner column on Forkable, I chatted with Kilo Code CEO Scott Breitenother about the project’s attempt to build an open source AI coding agent in a developer ecosystem shaped by closed platforms.

Kilo Code, fresh from an $8m funding round, helps developers plan, write, refactor, and debug software using AI. Its latest move highlights a growing tension in developer tooling: while open source has traditionally relied on shared code and community contribution, AI agents bring ongoing compute costs that don’t fit neatly with existing OSS funding models.

Read more: Forkable

Cloudflare acqui-hires Astro web framework team

Cloudflare has acquired the team behind the open source Astro web framework as part of its push to accelerate high-performance web development. The company says the Astro team will continue building tools focused on speed, simplicity, and modern web architecture, with Astro remaining open source.

Read more: Cloudflare

NetBird nabs $10m for open source network security

NetBird has raised $10 million in a Series A round of funding to expand development of its open source VPN and secure networking platform, positioned as a modern alternative to legacy enterprise VPNs.

Read more: Tech.eu

Signal creator turns attention to AI infrastructure

Moxie Marlinspike, best known as founder of open source messaging app Signal, is exploring how similar principles could be applied to AI systems.

As per an Ars Technica report this week, his focus with a new open source AI assistant dubbed Confer is on reducing centralization and opacity in AI tooling, arguing that without deliberate design choices, AI risks repeating the same trust and control failures seen in earlier platforms.

Read more: Ars Technica & Confer

Elon Musk says X algorithm will go open source

Elon Musk — not for the first time — has said that X plans to open source its recommendation algorithm, publishing the code that shapes what users see on the platform. The move is pitched as a transparency effort, though it remains unclear which components will actually be released, how usable the code will be in practice, or whether meaningful oversight will follow from making the algorithm public.

Read more: The Verge

Anthropic commits $1.5m to Python security

Claude Code creator Anthropic has committed $1.5 million to fund security improvements across the open source Python ecosystem, supporting maintainers and projects that underpin large parts of modern software and AI development. The funding is aimed at audits, tooling, and long-term maintenance, reflecting growing recognition that widely used open source dependencies remain under-resourced despite their critical role.

Read more: The New Stack

A warning on AI vs. open source economics

In a LinkedIn post this week that went somewhat viral, freelance engineer Marc J. Schmidt argues that AI coding agents are undermining the monetization models that have historically sustained popular open-source projects, forcing developers and maintainers to reconsider open distribution in favor of paid access or API-style gating. His core claim is that open source has long been funded by human attention (e.g. documentation visits, reputation, and expertise that convert users into customers) and that agent-driven development bypasses those attention funnels entirely, breaking the traditional economics of OSS.

The debate has been spotlighted by recent real-world developments in the web ecosystem: Tailwind Labs, the company behind the widely used Tailwind CSS framework, laid off about 75 % of its engineering team after revenue reportedly plunged around 80 % despite record usage, with its founder linking the decline to AI tools reducing traffic to documentation — the very channel that used to expose developers to paid products.

Read more: Marc J. Schmidt on LinkedIn

And finally…

A new lens on open source

Mentra, a startup building developer-focused wearable tech, has announced plans to start shipping its first smart glasses, Mentra Live, which are positioned as an open source alternative to the likes of Meta’s Ray-Bans. The device runs MentraOS, an open source operating system, and includes a built-in MiniApp Store and SDK that let developers build and install apps directly on the glasses.

Weighing around 43 grams, Mentra Live features a 12 MP camera for photos and HD livestreaming, along with support for calls, music, captions, and AI-powered features. The first limited batch of 1,000 units is set to ship from February 15, priced at $299, with further releases planned later this year.

Read more: Engadget & Mentra

In this edition, I check in with Scott Breitenother, CEO of Kilo Code, an open source AI coding agent that has introduced a new “momentum”-based subscription designed for AI-heavy development.

Kilo Pass reflects a growing tension in open source monetization, as AI-native projects introduce real, usage-based costs—from hosted inference to long-running agents—that don’t map cleanly onto flat subscriptions or traditional open core models. Kilo’s approach is to surface those economics directly, rather than hide them behind limits or access rules which ultimately shape how developers are able to use these tools.

“We didn’t want to play games with behaviour,” Breitenother told me. “Limits and throttling usually feel random and frustrating, especially when they interrupt you mid-flow.”

Read the story in full below.

Kilo Pass comes to pass

How do you price an AI coding tool when every keystroke has a real, recurring cost? That’s the million-token question Kilo Code is striving to answer with the launch of Kilo Pass, a new “momentum”-based subscription designed for regular, sustained use of AI in software development.

To rewind just a little, commercial open source startups (COSS) have long struggled with a familiar financial tension: how to build sustainable businesses around software that is, by design, freely available. The standard playbook typically revolves around things like hosted services, support contracts, or enterprise features layered on top of an “open core.” While there are pros and cons to these approaches, they all share a common characteristic: running the open source software itself typically does not impose significant, ongoing costs on the project’s developers, because execution happens on the user’s own infrastructure.

The rapid rise of AI adds a distinct economic wrinkle to that problem. Unlike traditional open source software, AI-powered tools carry direct, recurring costs every time they are used, driven by model inference that consumes GPU compute and electricity, and is typically metered in tokens—the units of text that AI model providers charge for processing. As developers increasingly rely on AI for longer-running tasks and autonomous agents, those costs scale with usage, but not always in ways developers can easily see or predict. For open source–aligned companies, this creates a dilemma: how to stay open and predictable without drifting toward limits, lock-ins, or licensing changes as usage grows.

It’s against this backdrop that Kilo Code, an open source AI coding agent, has launched Kilo Pass, a new subscription model built around what the company calls “momentum-based” usage. Rather than bundling AI access behind flat “unlimited” plans, Kilo converts subscription payments into non-expiring balances that developers spend directly on AI model usage, priced at the providers’ published rates, with additional bonus credits tied to sustained use.

The launch reflects a view that early pricing models across the AI coding landscape were misaligned with how these tools would actually be used in practice. Fixed monthly plans were simple to market, but they left little room for the sharp increases in compute demand that come with more frequent, longer-running AI-assisted workflows.

“A lot of AI coding tools started with flat subscriptions because everyone hoped model costs would drop fast enough to make unlimited use sustainable,” Kilo Code CEO Scott Breitenother explained to Forkable over email. “In reality, per-user usage scaled, and while the price of a specific model trends down over time, the cost of frontier models has remained stable.”

As those economics became harder to ignore, many products opted for incremental guardrails — adding caps, throttles, or slowdowns to rein in usage without revisiting the underlying pricing model.

And so Kilo’s approach speaks to a broader shift in how developers are responding to usage limits, model lock-ins, and opaque pricing as AI tools become more central to their work.

“That concern is valid, and we’re seeing it play out in real time,” Breitenother said. “Developers are pushing back on tools that combine limits with lock-in and unclear pricing.”

By way of example, Claude Code, Anthropic’s AI coding product, has become a recent flashpoint for developer frustration around limits and lock-in. While Anthropic offers higher-tier subscriptions aimed at power users, it recently restricted access to its Opus model when used through third-party tools.

As a result, developers paying for Claude’s higher-end plans could no longer use those benefits inside external coding environments such as OpenCode, Cursor, or Windsurf, and were instead confined to Anthropic’s own product surface. For many, that change underscored concerns that pricing and access controls were being used less to manage infrastructure costs, and more to shape where and how AI tools could be used.

“Kilo Pass is designed to go the opposite direction – we believe in being transparent about costs, not locking people into a single model or vendor, never throttling usage, and avoiding confusing credit systems,” Breitenother continued. “You can use a wide range of models, you pay true pass-through pricing, and the value in the pass comes from bonus credits, not restrictions.”

The path to Kilo Pass

Kilo launched some nine months back, courtesy of a founding team that includes GitLab co-founder and former CEO Sid Sijbrandij. Breitenother joined as CEO in September after the founding CEO stepped aside for personal reasons, and shortly after raised $8 million in seed funding.

At its core, Kilo Code is a multi-model coding environment designed to work across the IDE, command-line interface, and cloud-based agents. The project’s agent framework is open source, while model inference — the act of running AI models — is commercial, with developers paying for usage.

Until now, Kilo has charged developers directly for AI usage on a pay-as-you-go basis, with teams paying separately for team or collaboration features, and the option to bring their own API keys or run inference on their own infrastructure.

With Kilo Pass, the company is formalizing that usage-based approach into a recurring subscription, responding to growing frustration with how AI coding tools manage heavy usage behind the scenes, especially through pricing and access rules that shape day-to-day use.

“We didn’t want to play games with ‘behaviour’,” Breitenother said. “Limits and throttling usually feel random and frustrating, especially when they interrupt you mid-flow. Our principle from the start has been simple: don’t throttle people.”

Kilo Pass is offered as a monthly or annual subscription that adds money directly to a developer’s Kilo balance, which is then used to pay for AI model usage at providers’ list rates. Monthly plans are priced at $19, $49, or $199, with the subscription balance never expiring. On top of that balance, subscribers receive monthly bonus credits, which provide additional usage but expire at the end of each month.

Importantly, Kilo Pass doesn’t mark up model pricing or introduce usage controls. Instead, it’s structured to reward longer-term commitment, giving Kilo more predictable revenue while leaving developers in control of how they use the software.

“They’re [bonus credits] designed to reward people who are actively building, not to encourage stockpiling,” Breitenother explained. “The longer you stay subscribed, the higher your boost percentage becomes, up to a cap.”

Annual plans, on the other hand, cost the equivalent of twelve months paid upfront, but come with a key difference: subscribers receive a flat 50% bonus in boost credits each month, rather than gradually building toward that level over time.

In both cases, the bonus credits expire monthly, while the underlying paid balance remains available indefinitely.

When open source economics meets the reality of AI

Kilo’s move comes at a time when the broader open source ecosystem is grappling with how AI is reshaping sustainability. Just last week, the team behind Tailwind CSS – a widely used open source framework for styling web applications – cut the majority of its engineering staff after a reported drop in developer traffic. The project’s creator linked this in part to AI tools intermediating access to developer documentation, reducing direct traffic to the project’s site and their ability to convert developers to paid plans.

The episode underscored how AI is reshaping the relationship between open source projects and their users, altering where value is captured. In Tailwind’s case, AI tools increasingly sit between developers and documentation, weakening a key conversion path that had supported commercial products. Elsewhere, including at Kilo, the pressure shows up differently — through real, ongoing infrastructure costs tied directly to usage.

Against that backdrop, Kilo’s bet is that pricing models themselves need to evolve to reflect the real costs AI introduces, without undermining developer trust or openness. It also points to a broader tension now facing open source–aligned companies, as rising infrastructure costs collide with expectations of predictability and long-term sustainability.

Indeed, for Breitenother, Kilo Pass is perhaps as much about avoiding future trade-offs as solving today’s problems.

“A lot of licensing pressure comes from the same place,” he said. “Costs spike, pricing no longer matches reality, and companies reach for stronger control levers. Sometimes that shows up as throttling. Sometimes it’s lock-in. Sometimes it turns into licensing changes.”

If pricing can better reflect real usage, he argues, that pressure eases. “Kilo Pass is our attempt to prove you can keep trust high, keep the product open, and still support heavy real-world usage in a sustainable way.”

This week’s lead story looks at Google’s plan to slow the pace of Android source code drops to AOSP, and what that shift means for developers who depend on timely visibility into platform changes.

Elsewhere, there’s a tough illustration of open source sustainability pressures as Tailwind CSS cuts staff after an AI-driven traffic drop, plus a related note on the Open Source Pledge and the flow of funding commitments that followed. There’s also a security warning about VS Code forks, Microsoft open-sourcing its XAML Studio visual design tool, and more.

As usual, feel free to reach out to me with any questions, tips, corrections, or suggestions: forkable[at]pm.me.

Paul

Open issue

A new cadence for Android development

For some two decades, Google has published source code to the Android Open Source Project (AOSP) for every major and minor Android version, making each platform release available to the wider ecosystem. That approach gave developers, custom ROM builders, device manufacturers, and security researchers regular access to Android’s underlying changes, allowing them to sync repositories, build forks, and test updates as new versions arrived.

Starting in 2026, however, Google is reducing the frequency of those public AOSP source drops, moving from a roughly quarterly cadence to just two releases per year: once in Q2 and once in Q4.

Google says the shift aims to align with its internal “trunk stable” development model and improve platform stability for the broader ecosystem. Translated, this means simplifying Android’s internal development by working from a single codebase and releasing source code only at carefully controlled points.

First reported this week by Android Authority, the change is now visible on the official Android source site, where a new banner explains the policy and recommends developers build from the android-latest-release branch instead of the old aosp-main workflow.

For everyday Android users, this shift won’t affect how quickly security patches arrive on their devices, with Google confirming that monthly security fixes and security-branch updates will remain unchanged. But for communities that depend on consistently timely source access, the impact is more tangible.

Independent developers and projects such as custom ROM teams have already felt the effects of slower or delayed source availability. Last year, the source code for Android 16’s first quarterly platform update was notably late to land on AOSP, frustrating some in the community, though the following update arrived more predictably.

Android’s permissive Apache-licensed code has been one of the pillars enabling forks, alternative builds, and transparent platform development since the project’s early days. Reducing the cadence of public source drops doesn’t negate all of that, but it does change how and when developers gain visibility into Android’s changes, highlighting a familiar tension in vendor-led open source projects between control and transparency.

Read more: Android Authority

Patch notes

Tailwind CSS cuts staff after AI hits developer traffic

Tailwind CSS, a popular open source CSS framework that lets developers style websites using ready-made utility classes directly in HTML, has laid off 75% of its engineering team (from 4 to 1) after a reported 40% drop in website traffic. Project creator Adam Wathan attributed the decline largely to developers turning to AI tools instead of Tailwind help documentation, which had been a key driver for commercial signups.

The episode highlights how even widely used open source projects struggle to generate sufficient revenue for development, with AI worsening the crisis. Which leads us nicely to the next Patch notes story…

Read more: Search Engine Roundtable & GitHub

A strong tailwind for open source funding?

The issue of open source funding is a perennial challenge, but there are active efforts to address this. This includes the Open Source Pledge, an initiative launched by Sentry back in 2024 that commits companies to contributing a portion of their revenue to the open source projects they rely on — including another $750,000 annual allocation in 2025.

In the wake of the Tailwind news, the folks at Sentry, via its Syntax media brand, became a premium Tailwind sponsor to the tune of $5,000 / month. As did others from both the Open Source Pledge and elsewhere, including Google’s AI Studio team, as per this post from product lead Logan Kilpatrick.

The community stepping up to provide support is heartening, but it doesn’t change the reality that many open source projects remain financially fragile and reliant on goodwill, sponsorships, and other indirect revenue streams.

Read more: Sentry & Open Source Pledge & Adam Wathan (X)

VS Code forks expose users to malicious extension recommendations

Several forks of Microsoft’s open-source VS Code editor were found to be at risk of recommending malicious add-ons, after security researchers uncovered a flaw in how extension suggestions were handled.

According to security firm Koi, popular AI-focused VS Code forks such as Cursor, Windsurf, Google Antigravity, and Trae reused Microsoft’s built-in extension recommendation lists without fully adapting them to their own extension stores. As a result, these editors sometimes suggested extensions that didn’t actually exist in their marketplaces.

This opened the door for attackers to potentially upload malicious extensions using those names and have them appear as trusted recommendations inside the editor. Koi said it stepped in and claimed the unused extension names first to prevent abuse, and notified the affected projects.

While there’s no evidence that real attacks occurred, the issue highlights a broader risk in open source software: when projects copy shared components without also copying the security checks behind them, users can be exposed to unexpected threats.

Read more: Bleeping Computer & Koi

Microsoft ushers XAML Studio into the open

Microsoft open-sourced XAML Studio, a visual design tool used to build and preview XAML-based user interfaces for Windows apps. By releasing the project’s code publicly under the MIT license, Microsoft says it hopes to make the tool easier to extend, maintain, and adapt as Windows UI frameworks evolve.

Read more: Microsoft

Percona puts the ‘open’ in Everest

Database company Percona is set to rename its Kubernetes database platform Everest as OpenEverest, with plans afoot to move the project toward independent governance.

The company said the goal is to donate OpenEverest to the Cloud Native Computing Foundation (CNCF), while a newly formed company will take on commercial support and development responsibilities. The shift is meant to clarify Everest’s status as an open source project, reduce vendor control concerns, and give users clearer assurances about its long-term neutrality.

Read more: Percona

Bose has ‘sound’ open principles

As Bose prepares to wind down official support for its SoundTouch smart speaker platform, the company revealed it would make the platform’s technical specifications and APIs publicly available.

The move gives owners and developers the ability to maintain, replace, or extend the software used to control and integrate SoundTouch-enabled home theater and speaker systems after Bose ends updates. While the device firmware itself remains closed, the decision allows the community to preserve functionality and build alternatives to Bose’s official apps — a rare example of a consumer electronics company giving a second thought to users once a proprietary platform reaches end-of-life.

Read more: Ars Technica & Bose

Ghostty terminal gains nonprofit fiscal sponsor

Ghostty has secured nonprofit fiscal sponsorship, giving the project a formal legal and financial structure without becoming a company. The arrangement allows the open source terminal emulator to accept donations, manage funds, and pay for infrastructure while remaining community-focused and independent.

Read more: OMG! Ubuntu

Open Source Founders Summit returns for 2026

Ticket sales have opened for the 2026 edition of the Open Source Founders Summit, a two-day event aimed at leaders running companies built around open source software (read my feature interview with the summit’s creators from earlier this year). The summit focuses on the practical opportunities and risks that come with building commercial organisations on open source foundations, from funding and governance to growth and sustainability.

The 2026 event is set to take place in Paris on May 18–19.

Read more: Open Source Founders Summit & Tickets

And finally…

Scientists have a whale of a time

Researchers working with Project CETI (the Cetacean Translation Initiative) have published details of a new underwater system designed to record, analyse, and cautiously respond to sperm whale communication. Described in a recent peer-reviewed paper, the system focuses on distinctive click patterns — known as codas — that whales use for social interaction. Rather than attempting direct translation, the researchers analyse structure and repetition in these sounds and observe how whales respond when specific signals are played back.

What makes the work notable is its commitment to open source science. The team has openly released both the hardware and software used to capture, process, and analyse whale vocalisations, along with detailed designs and fabrication notes. This allows other researchers to reproduce the system, validate the results, or build on the work — turning what could have been a bespoke research tool into a shared scientific resource.

Read more: PLOS One & Earth.com