There are two ways to run agents on OpenComputer. This page covers hosted agents — platform-managed agent definitions you reach through persistent instances or ephemeral sessions, and connect to messaging channels or a public API. If you’d rather create a sandbox yourself and drive a Claude session inside it directly with the SDK, see the Agent SDK.
Cores
The AI runtime that powers your agent. OpenComputer ships two blessed cores:
Picking a core is the one decision that matters most — it determines the agent’s capabilities, config model, and extension surface.
Packages
Stateful extensions beyond the core — installed viaoc agent install. Packages can provision infrastructure (managed databases, vector stores) and wire it into the agent transparently.
| Package | What it adds |
|---|---|
gbrain | Persistent memory — 34 MCP tools for pages, vector search, graph, timeline, files. Backed by a managed Postgres database. |
Channels
Messaging platforms connected to your agent.| Channel | How it works |
|---|---|
telegram | Webhook-based. Messages flow through the OpenComputer gateway to your agent and back. |
Three commands to a running agent
Two ways to interact
Everything you can do with the CLI you can do with the REST API, and vice versa.CLI
oc agent commands — create, connect, install, inspectREST API
HTTP endpoints for agents, instances, sessions
Agents, instances, sessions
An agent is configuration. Its actual compute runs in one of two modes:
Managed agents (those with a core) get an instance automatically on creation. Raw agents — which have no core and use your own
snapshot and entrypoint — create instances or sessions explicitly via the REST API. Managed and raw agents coexist on the same platform.
Authentication
All requests require an OpenComputer API key viaX-API-Key header or --api-key CLI flag.
https://api.opencomputer.dev