Skip to main content
Give an agent a prompt and a task; OpenComputer runs it on a managed runtime and records every step as a durable, streamable, resumable session — stream it live, steer it with a message, and get results by webhook.

​
What the platform handles for you

Self-healing, managed runtime

  • Crash → auto-restart, picks up where it was
  • Idle → hibernate (≈ free); an idle session survives sandbox loss
  • You never restart anything — details

Brain / hands isolation

  • The brain (agent loop) runs apart from the hands sandbox (its files + commands)
  • Untrusted work stays contained
  • Your model key never enters either sandbox — it lives in the secret store

Durable log & resume

  • Every step in one ordered, durable log
  • Reconnect from any cursor — nothing missed

Reliable delivery

  • At-least-once, retried, dead-lettered
  • Signed (Standard Webhooks); dedupe on webhook-id

​
Get started

1

Define an agent (with your key)

A reusable { name, model, prompt, runtime } — POST /v3/agents. Pass your Anthropic key inline; it’s stored as a credential. (Agents)
2

Start a session

POST /v3/sessions { agent, input } — the agent starts immediately; you get the session and a browser-safe client_token.
3

Watch, steer, get notified

Stream GET …/events?stream=sse (resume from any cursor), POST …/messages to steer, and register a webhook for push.

​
How it works

Durable Agent Sessions architecture: your app starts/steers and streams a durable event-log session; a runtime (the brain) reads and appends to it and acts in a sandbox (the hands); the model runs on your key from the secret store, which never enters a sandbox; user-level events are delivered out via your webhook.
The event log is the durable spine: you start and steer against it, stream from it, and it drives the runtime. The brain (agent loop) is isolated from the hands sandbox (where commands and file edits run), your key lives in the secret store — never in a sandbox — and user-level events are delivered out to your webhook.

​
When to use it

Reach for sessions when you want a long-running or interactive agent and don’t want to build the plumbing above yourself. For a one-shot command in a raw box, use a sandbox directly.

​
Next steps

Quickstart

Start a session, stream it, and steer it in a few minutes.

How sessions work

How a session works — lifecycle and events.

Example apps

A repo of examples built on this API.