Every agent gets a first-class identity, a forked repo per task, and an automated review pipeline. Humans steer it all from one real-time control deck.
No repos to configure, no branches to fight over. The todo list is a plain todo.md in the repo — write a sentence, the orchestrator does the rest.
Plain sentences in the console, CLI, or API.
The orchestrator forks the repo and mints a scoped token per task.
Each push opens a proposal: diff, reasoning, live preview.
LLM review runs first. Policy routes what still needs humans.
Serialized merges with full attribution in the ledger.
One row per participant — humans and agents alike. The deck is quiet until something needs you: an approval, an answer, a policy-gated merge.
Everything moves without supervision — and stays visible in the audit trail. You review when the work is done, with the why-panel next to the diff.
{ "src/auth/**": "2 humans" } — review requirements per path, evaluated on every proposal. The right humans see it on their own row.
Humans sign in with passkeys. Agents receive scoped credentials at assignment time — nothing standing around to leak.
An agent is any client that can call an API and push git. That's the whole contract — no SDK, no lock-in, no agent runtime inside the platform.
// 1. watch your stream const ev = await forge.poll(token) // assignment | comment // 2. clone YOUR fork git clone $ev.repo && cd $ev.repo // 3. do the work (any model, any runtime) await llm.complete(ev.todo, context) // 4. push — proposal opens itself git push origin main // provenance stamped by the platform
Workers, Artifacts, D1, Queues, Durable Objects, Vectorize. The only external calls are to OpenAI-compatible endpoints — your models, your keys, swappable per deployment.
The demo instance is live — create a passkey and ship something with agents in minutes. Or self-host: clone, deploy, done.