Terminal

Fleet → Terminal gives an interactive shell on a server without SSH, a bastion, or an inbound port.

Requirements

Three things, all of which must hold:

1. The agent's profile is full.
2. You hold a grant for that server.
3. The grant has not expired.

Any one missing and the button does nothing useful — the page says which.

Authentication

A browser cannot set an Authorization header on a websocket. So the terminal uses a single-use ticket, minted by an ordinary authenticated POST /api/v1/agents/{id}/shell/ticket.

The HTTP fallback

The websocket depends on Connection and Upgrade surviving every proxy hop, and those are hop-by-hop headers that any hop may drop.

There is therefore a second transport: SSE out, POST in, over ordinary HTTP — the same primitives the rest of the application already runs on. It is ticket-authenticated in exactly the same way.

If the terminal works on the fallback and not on the socket, a proxy is stripping the upgrade. /api/v1/agent/ws-check will confirm it.

As whom

A session runs as the OS identity on the grant — root or a named user. It is resolved by the agent and capped by the machine, so a grant naming an identity the machine will not accept is refused there rather than honoured.

Sessions die with the grant

A grant that expires ends the session. Not at the next reconnect — immediately.

Recording

Sessions are recorded and audited: who, which server, which identity, when, and what was run.

Local secrets are redacted unconditionally — including in the transcript. An agent credential does not become readable because somebody cat-ed a file in a recorded session.

When you should not use it

The terminal is for investigation. For anything you do more than twice, use a declared command: it is reviewable, repeatable, works across many servers at once, and does not need full.