A programming language for AI agents.
In Tin, the compiler checks what an agent may do — which tools it can reach, what needs a person's yes, and where a model's words may go — before the program runs. It is for developers who ship agents, and for the coding agents that write programs for them.
$ curl -fsSL https://tin.ad/install.sh | bash
Learn Tin Deploy an agent Why a language
From install to a live agent in under five seconds
Install, write a four-line agent, and tin deploy it. Every agent becomes a streaming HTTPS endpoint at its own name.tin.ad, with WebSocket, MCP and an EIP-8004 identity file (an Ethereum proposal for agent identity) beside it.
$ cat hello.tin
/// Says hello to whoever just deployed it.
agent Hello {
model: "bedrock:us.anthropic.claude-haiku-4-5-20251001-v1:0"
system: "Answer in one short, friendly sentence."
}
$ tin deploy hello.tin --yes
uploading hello.tin (3 KB) to https://api.tin.ad
live: https://eager-wombat-a7tp.tin.ad anonymous, expires in 24 h (tin login keeps it)
$ curl -N https://eager-wombat-a7tp.tin.ad/agents/Hello \
-H 'accept: text/event-stream' -d '{"prompt":"hi"}'
event: text
data: {"text":"…"}
…
event: done
data: {"text":"…","turns":1,"tokens":…}
--yes answers yes to the one question a deploy asks: whether to send your model key from the environment to this deployment's secret store. Without it, you are asked. No key? model: "mock" deploys and streams the same way. Without an account a deploy lasts 24 hours, 10 per address per day; tin login keeps it.The compiler checks what an agent may do
Agent mistakes are not prompt problems; they are paths through a program. Tin finds those paths at compile time, so a tool that moves money two calls away from the agent still needs a person's approval:
error: tool 'pay' can run 'wire', which is guarded, without anyone approving it
--> sneaky.tin:4:35
= note: path: pay → payInvoice → wire
= fix: mark pay as guarded
- Injection doesn't compile. A request's input or a model's reply that reaches SQL, a shell or a URL is a compile error until it is parsed, matched or approved. Proof
- Capabilities are declared. A server lists what it may touch —
net.listen,model, files — and code that reaches past the list does not build. Proof - The deploy target is a capability. A voice agent sent to Vercel fails at build time with the reason and a fix, not in production. Proof
- Errors are written to be fixed. Each names the place, the path and the rule, and the safety errors say what to change, so the coding agent that wrote the program can repair it. Try it
Why now
Models increasingly write and run code, and every product is becoming an endpoint that an agent talks to. Today that code is glue across four languages, and nothing checks what the agent is allowed to do. The ground under one language finally exists: WASI 0.2, isolates at the edge, and small native binaries run the same program from browser to server to device.
What is measured today
| What | Result | Source |
|---|---|---|
| Install → streaming agent on tin.ad (Node 24+ present) | 4.6 s | tin.ad |
| Model providers written in Tin | 12; live on Claude Haiku 4.5 and gpt-oss-20b, the rest against recordings | providers |
| Native hello binary | 107 KB, 4.4 ms to run — on par with Rust (343 KB, 4.3 ms) | status |
| Recursive fib(36) on Node, median of 7 | Tin 53 ms · JS 99 · Python 667 · Rust 25 | bench |
| Where Tin loses | on Node: agent loop 112 ms vs JS 72; also hello, mandelbrot, nbody, trees | bench |
| Tin compiler written in Tin | byte-identical to the reference on 1,346 programs; rebuilds itself to the same bytes | status |
No users, customers or stars are claimed here. Everything still in progress is on /status.
Where it goes
Plan, not built yet: tin.ad becomes the place agents live — a signed name, a public identity through EIP-8004, and endpoints any other agent can find and call. The language makes those agents safe to let models write; the platform makes them one command to ship. If agents become a main way software is used, the place where they are named, checked and run is a business the size of web hosting. Read the memo.
Who builds it
Çağatay Çalı: on GitHub since 2014 with more than 450 public repositories, author of fastify/benchmarks, and working on agent frameworks and on agents that drive real robots. He directs coding agents in parallel lanes, and a critic agent that never edits the code must pass each one. About