Skip to main content
The console is where activity gets reviewed. It is served by airlockd itself — an embedded single-page app — at your deployment’s URL, and it is login-gated and multi-user from the first request.

​
Who sees what

Sign-in is email + password against Airlock’s own user table (httpOnly session cookie, server-side revocation). Two roles: Non-admins see the same timeline a reviewer would, scoped to themselves — enforced server-side, not in the UI.

​
Timeline

The core view is one chronological stream across all three event kinds:
  • Access events — proxy egress rows as compact ALLOW / DENY host chips, with method, path, byte counts, and duration.
  • Chat events — sandbox conversations: prompts, replies, tool calls and results, colour-coded, long content collapsed.
  • Decision events — an agent’s resolutions: target, decision, confidence, and cited sources, with any flags inline.
Filter by kind, user, and time range; pagination is keyset-based so deep history stays fast. Selecting a session opens the session drill-down — everything one agent session did, in order, across kinds.

​
The flags queue

Decision events run through the flagging rules the moment they are ingested, and flagged ones land in an admin review queue:
  • no_evidence — a decision that cites no sources.
  • low_confidence — self-reported confidence below your threshold.
  • disagreement — two decisions on the same target that do not agree.
Each flag opens with the decision’s evidence attached, such as sources, quotes, reasoning, and the session behind it; each flag also carries a status (open → acknowledged → resolved) with reviewer and resolution note.

​
User management

Admins create users, set roles, disable accounts, and mint push-ingestion tokens for external agent identities. They are named, scoped, revocable, and last-used visible. See deploy.