Each agent is a full Claude Code or Codex instance in its own isolated sandbox, with a shell, a persistent workspace, and git.
Talk to it from chat or drive it over the API. Close your laptop mid-task: the sandbox sleeps, wakes on your next message, and the work it packages as apps keeps running on schedules and webhooks.
The container stops when the thread goes idle and resumes on your next message with the same conversation, the same files, and the same signed-in agent. Your work stays put between sessions, so you can pick up exactly where you left off.
/workspace is a real volume shared across a project's threads. Repos, files, and data survive sleep, wake, and restarts.
Install the GitHub App and agents clone and push with short-lived tokens scoped to a single repo. No durable secret ever enters the sandbox.
Connect Claude or ChatGPT once per thread. The agent runs on your plan, with the real model catalog your account offers.
Every model call leaves through the Outgate gateway, so PII detection, routing, and observability apply to agent traffic too.
Describe a recurring job and the agent packages it as an app: a manifest, triggers, and a published page on its own URL. From then on the platform wakes the agent on schedule, so nobody needs to keep the tab open.
Chat is customer #0. It runs on the same public REST API you get. Create agents, stream their output, and wire webhooks from your own product.
Start a conversation in chat, or create your first agent over the API in a couple of curl calls. The harness handles the sandbox, the lifecycle, and everything the agent leaves running.