- Endless sessions. A session takes messages at any time, queues the ones that arrive during a turn, and processes them in order. It never ends on its own.
- Sleep and wake. An agent whose sessions are all idle sleeps, and the next message wakes it with its conversations and files exactly where they were.
- Recovery. A turn lost to a sandbox failure is run again over the same messages; a turn that stops making progress is ended by a watchdog and retried.
- Models included. Agents run on models paddock provides, with no provider account of your own, or on your own Anthropic or OpenAI providers and keys. No provider key ever enters a sandbox.
- Live state for your UI. Every message, tool call, and turn of a session is an event, streamed live and replayed after a reconnect, so your frontend writes no streaming, reconnection, or state code.
- Triggers. A webhook URL turns another service’s events, such as a GitHub pull request, into messages for an agent, with no server of yours in between.
How it fits together
Your backend publishes a template, a versioned YAML definition of a kind of agent. It creates agents from it, each with its own sandbox, and opens sessions inside them. It hands a session ID and a short-lived session token to your frontend, which shows the session live.Quickstart
Create an agent and chat with it from a browser in about five minutes.
Core concepts
Templates, agents, sessions, turns, events, and harnesses.
SDKs
Go, Python, and TypeScript on the server; TypeScript and React in the browser.
API reference
Every operation of the HTTP API.
Harnesses
paddock runs the harnesses unmodified, each at a pinned version:
You keep the mapping between your own users and paddock’s agents and sessions: store agent and session IDs next to your records, the way you store customer IDs from a payments provider.