Memo

One page for investors and partners. Measured facts are marked with their source; plans are marked as plans.

What are you building?

Tin, a programming language for AI agents, and tin.ad, the place Tin agents run. The compiler checks what an agent may do — reachable tools, required approvals, where model and request text may flow, what the deploy target supports — before the program runs. tin deploy turns a file into a streaming endpoint at name.tin.ad.

Why now?

Models write and run a growing share of code, products are becoming endpoints that agents call, and WebAssembly, WASI 0.2, edge isolates and small native binaries now let one language run in the browser, on servers and on devices. The missing piece is a language whose type checker understands agents.

What do you understand that others don't?

Agent safety is a property of program paths, not of prompts. SQL injection and prompt injection are the same bug — untrusted text reaching a powerful call — and one taint rule in the compiler stops both. Libraries see one call at run time; a compiler sees every path before deploy. And coding agents increasingly write the code, so the error message — what it names and what it says to change — is the product's real interface.

How far along are you?

The first commit was on 15 November 2026; the pace comes from the method below, not from an import — more than 440 commits since, each lane's work checked by a critic before it counts. Today the language, compiler, runtime, editor support, 12 model providers, WebAssembly, WASI 0.2 and native targets, a compiler written in Tin itself, and the tin.ad platform with GitHub sign-in all work and are tested. It is built by one person directing coding agents in parallel lanes, with a critic agent that must pass each lane. It is a preview: no public users yet, and the source repository is private while the preview lasts.

Numbers

MeasuredResultSource
Install → streaming agent on tin.ad (second machine, empty home directory, Node 24 present)4.6 s; first byte 0.44 stin.ad
Model providers, written in Tin12; live on Claude Haiku 4.5 and gpt-oss-20bproviders
Same output without JavaScript (WASI 0.2 and native)216 of 244 programs; 0 differ; 26 refused with a reasonstatus
Native hello binary107 KB, 4.4 ms to run; on par with Rust (343 KB, 4.3 ms)status
Tin compiler written in Tinidentical output on 1,346 programs; reaches a fixpointstatus
Speed vs JavaScript (Node 26), wall time incl. startupnative build faster on all 8 programs it runs; Tin on Node faster on 4 of 9, slower on hello, agent loop, mandelbrot, nbody, treesbench

Alternatives

How will you make money?

Plan. tin.ad is free during the preview. The business is hosting: paid plans for kept agents with higher limits, custom domains, teams, and usage-based compute, the way Vercel monetises Next.js. The language is free to use.

Risks

Write to Çağatay The thesis About