Admin page

Naming admins with ADMIN_EMAILS, what the gate accepts, and what an admin can see.

The portal has an Admin item in the sidebar, shown only to admins. It is the same portal, not a separate app.

Naming admins#

Set ADMIN_EMAILS to a comma-separated list and restart:

ADMIN_EMAILS=admin@example.com,ops@example.com

Matching is case-insensitive and ignores surrounding spaces. Empty, the default, means nobody is admin. There is no in-app promote or demote: the env var is the only source, so access is revoked by removing the email and restarting.

What the gate accepts#

/api/admin/* accepts the portal session only. An API key, including an admin's own MCP key, gets 401, so an agent never acts as admin. A signed-in user who is not on the list gets 403.

The email is read from the stored user record, not from the session token, and both SSO providers reject an unverified email, so an unverified address cannot reach the allowlist.

What the Overview shows#

Every section is its own tab with its own URL (/admin, /admin/activity, /admin/connections, /admin/custom-apps, /admin/profiles, /admin/vault, /admin/files, /admin/users, /admin/config), so the browser's back button moves between them. Each tab loads only when you open it, so nothing walks the profiles or the workspace on disk until you ask. They are read-only and fail independently.

The Overview tab holds:

• Headline numbers: tool calls and error rate over the last 24 hours (with the change against the 24 hours before), users active in the last 7 days, connections that need a reconnect, and disabled users. Usage figures need AUDIT_LOG_DEST=sqlite; otherwise they show a dash and say why, while the user and connection counts still show.
• Top tools: the most-called tools over 7 days with error counts and average time.
• Recent failures: the last five failed calls, with the error text, linking to Activity.
• Instance: version, database backend, cluster mode, audit-log destination, user count and the number of allowlisted admins (a number, never the list; it counts allowlist entries, whether or not they have signed in).
• Activity: every user's tool calls, newest first, with the user's email. Filter by app, errors only, or a user's email. An unrecognised status filter is refused with a 400 rather than ignored. Shows tool names and error text, never arguments. With AUDIT_LOG_DEST other than sqlite the events are not in the database, and the card says so.
• Connections: connected users per integration. "Needs reconnect" counts connections whose access token has expired and which have no refresh token; the server does not record refresh failures, so a connection that merely failed to refresh is not counted.
• Custom apps: every user's custom apps with owner and URL. Credentials are never shown.
• Browser profiles: disk used per user and whether the profile is live. "Live" means a use-marker file moved in the last hour. Workers and pods that share a profiles volume all show up here.
• Vault: counts only. Secrets stored, users holding any, secrets unused for 90 days, pending one-time links, and the users holding the most. No secret name or value is ever returned.
• Files: the agent file workspace. Disk used, file count, the age of the oldest file (older than WORKSPACE_TTL_HOURS means the reaper is not running), and the largest users and files by size. No file name or content is ever returned.

Managing users#

The Users tab lists every user with their connection and custom-app counts, last activity, whether they have an API key, and whether they are disabled. The search box filters by email as you type (case-insensitive, matching anywhere in the address). The list is capped at 500 users, and search covers only those; the tab says so when the list was cut off.

• Disable signs the user out of the portal and stops every credential: their API key, OAuth logins and refresh tokens, and new SSO sign-ins. Their connections, vault and files are kept, so Enable restores the account. Disabling asks for confirmation.
• Revoke key clears the user's API key; agents using it stop working until they create a new one.
• You cannot disable yourself or any email on ADMIN_EMAILS. Admins are managed by that variable, not here, and this also stops an admin locking everyone out.
• Every action is written to the audit log with your account as the actor.

Instance config#

The Config tab holds two instance-wide settings:

• Integrations: turn an integration off for every user. It disappears from tool search, schema lookup and execution, the REST endpoint and the connect flow. Existing connections are kept, and turning it back on restores everything.
• Custom apps: choose who can add custom apps: everyone, no one, or selected users. Excluded users keep their existing apps, but agents stop seeing those apps' tools.

Changes apply at once in the process that handled the request. Other workers and pods pick them up within INSTANCE_SETTINGS_POLL_SECONDS (default 5), so allow that long before assuming a change is everywhere. OAuth client secrets are not editable here; they stay in environment variables.