A coding model that answers to no one but you.

The most capable coding models on earth are locked behind login walls, opaque pricing, usage policies that change without notice, and the constant possibility of being de-platformed mid-project.

We think that's the wrong shape for something as fundamental as the ability to write software. The tools that build everything else should not themselves be enclosed. StrataCode exists to make open-source coding models genuinely accessible — not "open" in a press release, but open in the only way that counts: you can reach the model directly, on your terms, without asking anyone's permission.

So we did the unglamorous work. We built two coding agents tuned specifically for everyday software work, and put them on a peer-to-peer network where anyone can connect directly. There is no StrataCode dashboard to sign up for. There is no API key to be revoked. There is a peer ID, a price per token, and a node that answers when you call it.

Payment runs on USDC over Base, metered token-by-token through AntSeed's payment channels. You pre-deposit, requests draw down your balance, and providers settle on-chain. You are never billed for a subscription you forgot to cancel, and we never see a credit card, an email address, or a single line of the code you write.

This is a small node with a stubborn premise: that the right to compute, to generate, and to build should be distributed — held by the people doing the building, not rationed out from a console you don't control.

Both agents are deliberately lean — specialized, not sprawling. That's the point: they're cheap to serve, fast to first token, and tuned to fit exactly where the giants can't.

A generalist is mediocre at code, so we specialised hard. We assembled a corpus drawn exclusively from permissively-licensed public repositories — source, tests, diffs, issue threads, and documentation — and filtered for correctness, removing anything that wouldn't compile, pass its own tests, or carry a license we could honor.

On top of that we ran instruction tuning across the five categories the node advertises: chat, coding, structured JSON, step-by-step reasoning, and tool use. Each category got its own curated set of demonstrations — real refactors, real bug fixes, schemas that actually validate, function-call traces that actually parse — so each agent learns the shape of useful answers, not just the vocabulary.

We held back a private evaluation suite of held-out coding tasks and graded every checkpoint on whether the code ran and did the right thing, not on token overlap. StrataCodingAgent is the checkpoint that wrote the most working code per dollar; StrataCodingAgent-Fast is the lighter, lower-latency variant for when speed and throughput matter more than the last few points of accuracy.

Neither agent is a frontier model and neither pretends to be. They're focused, honest tools — strong at the dense, repetitive 80% of programming work: writing functions, generating valid JSON, drafting tests, explaining a stack trace, calling a tool. For that work, served at this price, with this much autonomy, we think they're hard to beat.