The Demand Curve That Files Itself

On Monday, someone noticed that Anthropic had shipped source maps in a Claude Code npm release. By Tuesday morning, the full source — nearly 2,000 files, half a million lines of TypeScript — was mirrored across GitHub. A post sharing the link hit 29 million views on X. A rewritten version of the repository briefly became GitHub's fastest-ever download.

Anthropic filed a DMCA takedown. GitHub processed it against 8,100 repositories, accidentally catching legitimate forks of Anthropic's own public Claude Code repo in the blast radius. Boris Cherny, Anthropic's head of Claude Code, called it unintentional and retracted the bulk of the notices.

That's the story everyone is reporting. Leak, takedown, oops.

The more interesting story is what happened next, and what it means when you zoom out.

The Rewrite Reflex

Within hours of the leak, developers weren't just mirroring the source. They were porting it to Rust.

Not because Rust is inherently better for this use case. Because Claude Code is closed-source TypeScript, and a cohort of developers looked at the internals (the prompt structures, the tool-routing patterns, the permission system, the context management) and decided the fastest path to something they could own and modify was to rewrite it in a language they trusted more.

The impulse is older than this leak. Neovim from Vim. LibreOffice from OpenOffice. Signal from TextSecure. When developers encounter a tool they depend on but can't control, some subset will always attempt to reproduce it under a license they can.

What's new is the competitive context. OpenAI's Codex CLI is open source. Google's Gemini CLI is open source. OpenCode is open source. Claude Code isn't. And in a market where the harness layer is commoditizing fast, keeping your harness closed while your competitors open theirs creates a specific kind of pressure.

The DMCA notice doesn't fix that pressure. It addresses the specific infringement. It does nothing about the Rust port, the Python reimplementation, the GPT-compatible fork, or the dozen other derivative projects that used the leaked architecture as a blueprint but rewrote the code. Those are technically clean. They're also direct competitors that wouldn't exist yet without the leak.

Every takedown that removes a mirror but leaves a rewrite is Anthropic fighting the symptom while the cause compounds.

The Corpus Nobody Reads

The Anthropic takedown landed in a specific place: github.com/github/dmca, the public repository where GitHub archives every copyright takedown notice it processes.

It's one of the most underread public datasets on the internet.

I spent an afternoon scrolling through the March 2026 notices. Not just the Anthropic one. All of them. A single month's filings:

Proprietary source code and internal tooling. Anthropic, Mercedes-Benz, Intel, Robinhood, PowerSchool, Nexteer Automotive, Autoliv, Synopsys. Companies discovering their internal code on public repos — sometimes leaked by employees, sometimes by contractors, sometimes by AI agents trained on proprietary codebases.

Developer tools and license bypasses. JetBrains IDE cracks. Emby media server unlocks. Lightspeed Systems school network management. These are professional-grade tools that developers use daily, and some subset refuses to pay at the listed price.

Premium creative assets. RPG Cameras & Controllers, a Unity asset generating repeated takedowns, alongside Pixelcave templates and game mods. The kind of things that save a game developer forty hours of work.

Academic and research materials. Springer Nature, academic journal articles, English language textbooks. Knowledge behind paywalls, reposted by people who apparently need it badly enough to host it on GitHub.

Sports streaming and media. La Liga, Premier League, VIZ Media manga. The traditional piracy category — but now sitting alongside JetBrains cracks and Mercedes source code in the same archive.

Laid out in sequence, the corpus stops looking like a legal ledger and starts looking like something else: a revealed-preference map of what people want badly enough to route around legitimacy.

Professional Leverage Goods

The interesting observation isn't that people pirate things. That's been true since Napster. The interesting observation is what category of things shows up when you filter the DMCA repo by developer-adjacent activity.

It's not mostly entertainment. It's mostly leverage.

A coding font makes you faster. A JetBrains IDE makes you faster. A Unity asset pack saves you a sprint. A textbook teaches you something your employer won't pay to teach you. Claude Code — the full source, not the minified distribution — gives you the ability to understand and modify the tool that writes your code.

These are professional leverage goods: artifacts whose primary value is making the person who holds them more capable, more productive, or more autonomous in their work. When a DMCA notice targets one of these, it's not the same economic event as taking down a movie torrent. The person mirroring a movie wants entertainment. The person mirroring Claude Code wants agency.

I don't think this is an argument for "everything should be open." I don't think Anthropic was wrong to file the takedown. I don't think the Rust porters are necessarily heroes.

But I do think the velocity of the response — 8,100 repos in hours, Rust rewrites by the next morning, GPT-compatible forks within days — is telling you something the DMCA notice itself can't capture. The demand for agency over your own development tools is high enough that thousands of developers will route around the legitimate channel the moment they see an opening.

That's not a piracy story. That's a pricing and access story wearing piracy's clothes.

What the Archive Actually Shows

Every DMCA notice is, in a narrow sense, a failed transaction.

Not failed as in fraudulent. Failed as in: someone wanted this thing, the legitimate path didn't work for them, and they found another path. Too expensive, not available in their country, gated behind an enterprise contract they can't sign, or simply not open-source when they believe it should be. The DMCA notice records the enforcement. It doesn't record the demand that preceded it.

But in aggregate, the patterns are legible. When the same Unity asset generates ten takedowns in a month, that's not ten crimes. It's ten data points showing that the asset's distribution model is fighting its own demand curve. When a coding font generates twenty-one takedowns, someone in marketing should be asking why their pricing page isn't converting the people who clearly want the product.

The Claude Code takedown is the most dramatic recent entry, but it fits a trendline. Developer tools are migrating toward open-source harnesses with proprietary model backends — the pattern I wrote about in March. Codex is open. Gemini CLI is open. The harness is the commodity layer. The model is the moat. Anthropic structured Claude Code the other way: closed harness, model behind it. The leak tested that structure against actual developer behavior.

The market's response was not ambiguous.

Wedges

The part that actually matters about the Claude Code situation isn't the source code.

The rewrites are potential wedge products. Rust, Python, compatible with GPT and Llama and local models. Each one starts as a novelty project built on leaked architecture. Some will die. Some won't. The ones that survive will attract contributors who have opinions about how a coding agent should work, and those opinions will diverge from Anthropic's product decisions.

This is how wedge competition forms. Not from a startup raising money to compete with Claude Code. From a maintainer who forked the architecture, rewrote it in Rust, made it work with three different model providers, and gradually attracted a community that prefers the open version.

Is it wrong? I genuinely don't know. Should it be illegal? The rewrites are probably clean — copyright protects expression, not ideas, and a clean-room reimplementation of architecture patterns is standard practice.

But legal or not, it creates real problems for Anthropic that compound over time. Every clean-room rewrite that gains traction is a customer who would have been locked into Claude Code's ecosystem and now isn't. Every multi-model fork that works well enough is evidence that the harness layer isn't where the value accrues.

The DMCA notice removed 8,100 mirrors. It didn't — couldn't — remove the architectural knowledge that's now distributed across thousands of developers' heads. You can take down the repo. You can't take down the understanding.

The Demand Signal

The Claude Code leak will get processed as a security story, a legal story, maybe a "this is bad for Anthropic's IPO" story. All true. All boring.

The more durable observation is simpler: GitHub's DMCA repository is an accidental public record of what developers want badly enough to route around the rules, and the composition of that record has shifted. It used to be mostly entertainment. Now it's increasingly leverage — tools, datasets, source code, assets, and agent infrastructure.

Every one of those takedowns is a demand signal from someone the market didn't serve. Some of those people are freeloaders. Some are in countries where the pricing doesn't make sense. Some are building something and need a component that's paywalled at enterprise rates. Some are developers who want to own their tools.

The interesting part of the Claude Code leak isn't that Anthropic lost control of a repository. It's that 8,100 copies appeared in hours, rewrites appeared the next day, and the derivative projects that emerged are technically clean enough to survive the takedown.

That's not a leak story. It's a demand curve — and it filed itself.