Status
Every feature of Tin, sorted by what you can rely on now, with the date it last changed and how it was checked.
Works 20 · In progress 0 · Planned 6 · Providers
Works
You can run or check these today. The home page shows only these.
-
Agents, tools and guarded tools
agent, tool fn and guarded are keywords; the tool JSON comes from your types; a path to a guarded tool that skips a person does not compile.
Checked: the home page example and its compile error are produced by the real compiler on every docs build · Home · Learn: first agent
-
Real models, verified live
Bedrock Claude Haiku 4.5, and gpt-oss-20b through an OpenAI-compatible endpoint. Every live case is also recorded to a tape and replayed offline.
Checked: npm run test:live (tests/live cases 1–6 and 9) green on both models · Providers · Learn: real models
-
Other model providers
Anthropic, OpenAI (Chat and Responses), Gemini, Mistral, Ollama, llama.cpp, Llama API, LiteLLM, Writer, Vercel AI Gateway, plus "a | b" fallback and routing. Implemented and tested against recordings, not yet against each live service.
Checked: one provider test file each (tests/*-provider.test.ts), scripted HTTP · Providers
-
Anthropic API parity in the agent
Every item of @anthropic-ai/sdk 1.8.0 that a model call uses is supported or marked not applicable with its reason: thinking (adaptive, budget, between tools, interleaved, display), effort, stop sequences, tool choice, caching with a 1-hour TTL, headers and betas (headers and body on every provider), structured outputs, images, PDFs, documents with citations, search results, every stop reason, usage, errors. Claude on Bedrock speaks the Messages API with bedrock-anthropic:.
Checked: docs/parity/anthropic.json, generated from the SDK types; npm test fails on a missing or unlisted item; tests/agent-opts.test.ts; 61 items checked live on Claude Haiku 4.5 through Bedrock, recorded to tests/tapes/aparity-*.json and replayed · Parity matrix · Agent options
-
Tin in the browser
The compiler and your program run in the tab as WebAssembly. Agents in a page need Chrome or Edge 137 or newer (JSPI).
Checked: every Learn snippet runs in a headless browser on each docs build (learn-play check) · Playground
-
Record and replay
tin run --record writes model replies, fetches, input and random numbers to a tape; --replay runs from it with no model and no network.
Checked: npm run test:live records every live case to a tape and replays it offline · Learn: replay
-
Capability report and taint
tin caps lists what each agent can reach before it runs; model output cannot reach outside code without approve.
Checked: tests/taint.test.ts; the tin caps report on Learn: capabilities · Learn: capabilities
-
Editor and packages
tin lsp (errors, hovers, rename, completion, format), a VS Code extension in the source, tin add with tin.toml and tin.lock, tin doc.
Checked: LSP and package tests in npm test · The tin command
-
Benchmarks
The same programs in Tin, JavaScript, Python and Rust, 7 runs each, with the method and every caveat.
Checked: bench JSON shipped with the site; rerun with npm run bench · Benchmarks
-
Model providers written in Tin
All 12 text providers and routing are Tin code (std/models), including SigV4 signing and the Bedrock event stream, which is how native agents reach real models. Live through the Tin code: Bedrock (Haiku 4.5, Nova Lite), gpt-oss-20b via an OpenAI-compatible endpoint, and the Responses API; Gemini, Ollama and Llama API are checked byte-for-byte against recordings only. Voice providers stay in the host.
Checked: provider tests unchanged; npm run smoke:bedrock 6/6; request bodies byte-identical to the TypeScript providers
-
A Tin compiler written in Tin: tinc
compiler/*.tin (about 15,000 lines) is a second Tin compiler and a command: tinc build app.tin -o app.wasm and tinc check, with imports read from disk. It writes the same bytes as tin build and the same errors as tin check on every program in the test corpus and every file in examples/, and compiles itself to a fixpoint. As a native binary it needs no Node and no wasm runtime. It writes modules for the JavaScript host; WASI and native builds still go through tin build.
Checked: npm run parity:tinc: 634 of 634 builds and checks identical (253 byte-identical builds); tests/tinc.test.ts (fixpoint through the CLI on files, native tinc); tests/selfhost-*.test.ts (checker and code generator 100% of the corpus, 0 mismatches) · tinc
-
Files
readFile, writeFile, readBytes, writeBytes, exists, isDir and listDir, behind the fs.read and fs.write capabilities (tin caps and server grants show them). Under tin run, WASI (only the --dir directories) and native, with the same results and error texts; in a browser they are an Err. A request may not choose a file path, and a model reply may not choose the path written without approve.
Checked: tests/fs.test.ts: one program on Node, wasmtime (0.2 and preview 1), the Host ABI reference and a native binary, identical · tin build --target wasi
-
tin deploy to tin.ad
curl -fsSL https://tin.ad/install.sh | bash installs tin (Node 24 or newer), then tin deploy app.tin gives a <name>.tin.ad address that streams each agent over Server-Sent Events, with an EIP-8004 registration file. Without signing in a deployment lasts 24 hours; tin login (GitHub) keeps it, and tin ls, logs, rm and secrets work from any machine.
Checked: cloud/three-minute.sh on a second machine (Mac mini, empty HOME, Node 24.22.3, 14:33Z): install, deploy and a streamed Claude Haiku 4.5 answer in 4.6 s, first byte 0.44 s; cloud/login-smoke.sh for sign-in; tests/cloud*.test.ts · tin.ad · Deploying
-
Servers and request taint
A server declaration with routes, WebSockets and net.listen as a capability. Request input is tainted like model output: it cannot reach outside code or a fetch URL until parsed or encoded. Each request runs in a fresh instance, so a panic answers 500 for that request only.
Checked: bench/load.mjs (bench/results/load.json): tin serve, mixed routes with a panicking one among them, 60 s at 64 and 256 connections and 180 s at 64: 979,536 requests, 0 wrong answers, p99 116 ms at 64 connections, memory flat (238 MB peak in the first quarter, 162 MB in the last); live on a tin.ad deployment at 64 connections: 1,000 answers, 0 wrong, then the anonymous daily cap. tests/server.test.ts (grants, request taint, sanitizers); review closed with 0 blockers (TINAD-CRITIQUE.md) · Learn: servers · Servers
-
Build targets for clouds
tin build --target cloudflare | vercel | cloudrun | azure | lambda | pages. A feature the target cannot run (WebSockets on Vercel, for example) is a build error. Cloudflare and Vercel are deployed live. Cloud Run, Azure Functions and Lambda builds run under each platform’s own local runtime, with a real model; they have not been deployed to those clouds (see Planned).
Checked: live demos on Workers and Vercel; a Bedrock agent on a Workers build streams under workerd (wrangler dev, 01aa4d7); scripts/emulate-targets.sh: Claude Haiku 4.5 streamed through the Cloud Run container command, the AWS Node runtime client (aws-lambda-ric 4.0.2, response streaming) and Azure Functions Core Tools 4.16.1; tests/cloud-targets.test.ts and tests/deploy*.test.ts · Deploying
-
WASI 0.2, no JavaScript
tin build --target wasi makes a WASI 0.2 component; fetch() and http() work under wasmtime run -S http, spawn, within and sleep run on a scheduler in the module with no JavaScript threads, approve asks on the terminal, and agents run (the mock model in the tests; real models through the providers written in Tin, checked by an opt-in live test), A.save / restore included. A server runs under wasmtime serve with its agent endpoints (GET /, /agents/<Name>, /mcp); a file with agents and no main becomes their server, as with tin serve. An event stream from an agent endpoint arrives whole, when the answer is done. Not on WASI, by design: WebSocket routes (wasi:http 0.2 has no connection upgrade, so a ws route is a build error), voice (audio devices and a realtime session over a WebSocket), import js / py / mcp (they call a JavaScript or Python host) and code: true (it needs the compiler and a wasm engine inside the module).
Checked: tests/wasi-p2.test.ts and wasi-serve.test.ts: output identical to tin run and tin serve; tests/nojs-endpoints.test.ts: every agent-endpoint and MCP request answers under wasmtime serve as under tin serve; examples/agents/12-resume.tin (save / restore) prints the same under wasmtime as under tin run · Install: where Tin runs · tin build --target wasi
-
Native executables
tin build --native makes one executable with no runtime to install. fetch() and http() work (with TLS), spawn runs without threads, approve asks on the terminal, agents (tools, typed replies, A.try, A.save / restore, agents as tools) answer from real models (Bedrock, Anthropic, OpenAI-compatible), and a server listens with its agent endpoints and /mcp. --cross linux-arm64 and linux-x64 build static Linux executables, run on Linux with the same output. Not native, by design: voice, import js / py / mcp and code: true (as on WASI; code: true is a build error that says so). WebSocket routes are not native yet (see Planned).
Checked: tests/native*.test.ts and tests/nojs-endpoints.test.ts (routes and agent endpoints identical to tin serve); 12-resume.tin identical to tin run; npm run smoke:native: Claude Haiku 4.5 on Bedrock answered live with an empty environment except the key; npm run parity:linux (16:15Z, Ubuntu 26.10, kernel 7.3 aarch64, Lima VM; linux-x64 under qemu-x86_64): 244 programs, 427 builds identical to tin run, 0 differ, 57 refused at build time with the reason; examples/agents/hello-native.tin answered live from Bedrock as linux-arm64 and linux-x64 · tin build --native
-
Voice agents
voice agent holds a speech-to-speech session with barge-in, tools mid-call and typed events. Live on OpenAI Realtime: in tin run, and as a deployed tin.ad agent at wss://<name>.tin.ad/agents/<Name>/ws, where a client sends PCM16 audio and gets the spoken reply back. Gemini Live and Nova Sonic are implemented and pass against the mock; they have not answered live (see Planned).
Checked: npm run smoke:bidi on gpt-realtime-2.1-mini: a recorded question transcribed, the weather tool called mid-call, a spoken answer using it, and barge-in, all pass; bench/voice/ws-client.mjs against a tin.ad deployment (bench/results/voice-live.json): 3 of 3 calls answered with the tool result, first reply audio 0.96–1.29 s after the question ended, 3.0–6.1 s of speech back; tests/voice.test.ts, tests/bidi.test.ts, tests/openai-realtime.test.ts · Voice agents · Learn: voice
-
Robots and policies
robot Arm = so101 from a registry of 7 robots; joint names checked at compile time; angles as units; arm.run(policy, 5s, 50Hz) with holds, sequence, chunked, perturb and replay; one safety gate for every write (ranges, safe bounds, speed, validators, estop); arm.drive lets an agent move the arm with every step refused outside the bounds and one approval per drive; a model reply reaching act/reach/run does not compile; run tapes replay exactly. Worlds: the built-in kinematic one (no physics; same trajectory under tin run, wasmtime and native) MuJoCo physics (Google DeepMind’s WebAssembly build) under tin run, live in the playground and linked statically into native executables with the same digits; and world: py, the robot in strands_robots through Python. Tin’s MuJoCo and strands_robots agree to 2.1e-9 rad on the same model.
Checked: tests/robot-*.test.ts (decl, kin, run, gate, drive with a scripted model, taint, tape, world incl. the SO-101 in MuJoCo when its files are cached), examples/robots/* identical under tin run, wasmtime and as a native executable, examples/mujoco/wave identical under tin run, Chromium and native (tests/robot-mujoco-run.test.ts), tests/robot-compare.test.ts (strands_robots parity), examples/learn/32-34 · Robots · Learn: robots
-
Learned robot policies
onnx("holosoma?vx=0.5") runs an ONNX policy in the control loop, behind the same gate: the reference presets (holosoma on the G1; microduck named), any .onnx file, hub("owner/repo") with each file checked against the Hub SHA-256 and an allowlist (TIN_POLICY_REPOS) for repos named at run time, LeRobot (PyTorch) checkpoints through world: py (strands_robots loads them, checks their camera and state requirements, Tin’s gate writes each action), and remote("https://…" | "wss://…") policy servers whose actions are refused, not clipped, when out of bounds. Literal specs are checked at compile time against the robot and its world. onnxruntime-node under tin run, onnxruntime-web in the playground; weights downloaded once at a pinned commit and SHA-256, never bundled.
Checked: tests/robot-policy.test.ts; tests/robot-policy-compare.test.ts: 500 ticks fed to strands_robots’ HolosomaPolicy, actions bit-identical; the G1 walks 3.87 m in 10 s under tin run (the reference: 3.87 m in its own loop) and 3.2 m in 8 s in Chromium (docs/scripts/check-policy-play.mjs); hub("lerobot/smolvla_base") ran 10 ticks on the SO-101 in world: py (TIN_TEST_LEROBOT_PYTHON) · Learned policies · Watch it walk
In progress
Nothing in progress.
Planned
Not built yet.
-
Robots: hardware and cameras
Real arms (the SO-101 over its serial bus or through Python), camera images in a Tin Observation (a LeRobot checkpoint in world: py sees them through strands_robots), runs paced to the wall clock with missed ticks and a watchdog, learned policies in WASI and native builds, and agents calling robot tools mid-turn (they drive through arm.drive instead).
Checked: listed under Not supported yet on /ref/robot
-
Voice on Gemini Live and Nova Sonic, live
A spoken reply from Gemini Live and from Nova Sonic. Both providers are built and tested against the mock.
Checked: needs a Gemini API key that is not reported as leaked, and AWS access keys for Nova Sonic (a Bedrock API key is refused) (owner)
-
Left for hosts without JavaScript
WebSocket routes in a native server (the native listener takes HTTP/1.1 requests only, no upgrade yet), spawn inside a server, tool parameters of type i64 or an enum with data, and telemetry spans. Each is a build error that says so; tin run, tin serve and tin deploy run them.
Checked: src/nojs-info.ts, checked against what the builds refuse by tests/nojs-info.test.ts
-
Live deploys to Cloud Run, Azure and Lambda
tin deploy --target cloudrun | azure | lambda against a real account. The builds run locally under each platform’s runtime; a live deploy needs a Google Cloud project, an Azure subscription or an AWS account, and none is set up yet.
Checked: needs an account (owner)
-
tin itself in Tin
tinc doing what tin build --target wasi / --native and tin run do (linking the runtime, wasm2c), and packages from git, so the whole tool needs no Node.
Checked: tracked in the tinc plan
-
A public package and source
tin on npm and the source repository opened.
Checked: see Install
Model providers, one by one
Verified live means a real model answered a Tin program through that provider. Every provider also has offline
tests against recorded or scripted replies in npm test.
Text models
| Provider | Model name | State | How it was checked |
|---|---|---|---|
| Bedrock | bedrock:…claude-haiku-4-5 | Verified live | test:live cases 1–6 and 9 pass; smoke:bedrock 6/6 (2026-11-16) |
| Bedrock | bedrock:…nova-lite | Verified live | smoke:bedrock: the agent + tool program and prompt caching; routing case of test:live |
| OpenAI-compatible endpoint | openai:gpt-oss-20b + OPENAI_BASE_URL | Verified live | test:live cases 1–6 and 9 pass, 54/54 agent examples; also through the provider written in Tin |
| Bedrock, Anthropic Messages API | bedrock-anthropic:…claude-haiku-4-5 | Verified live | thinking, interleaved thinking, stop sequences, custom headers and a tool round trip, live; recorded to tests/tapes/aparity-thinking-tool.json (2026-11-16) |
| OpenAI | openai:… | Reached, no reply yet | reaches api.openai.com and authenticates; no reply yet, because the key used had no credits (429). The same code path is live through the OpenAI-compatible row. |
| OpenAI Responses | openai-responses:… | Recordings and mocks only | not run live (same key) |
| Anthropic | anthropic:… | Recordings and mocks only | no key was available; Claude is live through Bedrock |
| Gemini | gemini:… | Recordings and mocks only | no working key |
| Mistral | mistral:… | Recordings and mocks only | no key |
| Ollama | ollama:… | Recordings and mocks only | no local server |
| llama.cpp, Llama API, LiteLLM, Writer, Vercel AI Gateway | … | Recordings and mocks only | no keys or servers |
Voice models
| Provider | Model name | State | How it was checked |
|---|---|---|---|
| OpenAI Realtime | openai-realtime:gpt-realtime-2.1-mini | Verified live | npm run smoke:bidi: transcribed question, tool mid-call, spoken answer, barge-in; 3 of 3 calls answered on a tin.ad deployment (bench/results/voice-live.json, 2026-11-16) |
| Gemini Live | gemini-live:… | Recordings and mocks only | fixtures only; no working key |
| Nova Sonic | bedrock-sonic:amazon.nova-sonic-v1:0 | Recordings and mocks only | fixtures; request signatures match the AWS SDK signer byte for byte; it needs access keys, which were not available |
| mock | mock | Works offline (not a real model) | whole sessions end to end, offline: tests/voice.test.ts, the playground with a fake microphone |
The source is private for now, so test files are named rather than linked. Built from c9137d6, Tin 0.1.0.