Odyshell documentation

Odyshell is infrastructure for external AI Agents that need temporary shell authority on private Linux Machines. Agents connect through remote OAuth MCP or canonical HTTP. Humans configure trust and policy in the dashboard and may supervise Tasks, but are not required in the execution loop.

The path to a Command

1. An Organization owns the Agent, Machine, policies, Task, and audit trail.
2. A registered Agent requests one temporary Task for one Machine.
3. Local Policy sets the Machine owner's non-bypassable ceiling.
4. Autonomy Policy autoapproves a matching Task or routes it to an Owner, Admin, or Supervisor.
5. The Agent submits asynchronous Commands and polls bounded output by cursor.
6. The Agent explicitly completes the Task after every Command is terminal.

The Linux Client connects outbound. It exposes no inbound port and shares no SSH credential or private-network route. Commands run as the Client's operating-system user with no sandbox or sudo setup, so use a dedicated least-privilege user.

Start with the Quickstart, then read Machines, Agents, MCP, and Self-hosting.