Add an app
Add an app from the catalog, or add a custom MCP endpoint:From the catalog
Choose a reviewed managed API integration or a probe-verified MCP connector. Members click Connect; Passport applies
the workspace’s catalog policy automatically.
Add custom app
Point Passport at any MCP server URL. It connects, discovers the tools, and scans them.
org), each person’s own account (member, through an OAuth broker), or no auth (none). Managed API integrations use per-person OAuth.
Some connectors need a one-time company OAuth client before anyone can connect them (when the provider does not
support dynamic registration). Passport shows a “needs admin setup” state and the callback URL to register, rather
than a dead-end member sign-in.
Enterprise catalog modes
- Open: verified catalog apps connect without an admin decision.
- Review: the first Connect intent creates or joins a grouped request.
- Rules: allow or review by app, category, and team, with an explicit block for exceptions.
Passes and access
A pass is what a person can do with an app:- Full: every tool, including writes.
- Read-only: only the tools marked read-only. Write tools are hidden from the tool list entirely, so an agent never sees a capability its user cannot call.
- None: no access; the app is not advertised to that person.
1
Workspace default or selected scope
Grant a default pass to everyone, or scope the app to selected teams and people only.
2
Per-team passes
Give a team its own pass, for example read-only GitHub for Support.
3
Per-person overrides
Override a specific person, which wins over the team and default.
How read-only is decided: a tool is treated as read-only only when it is provably so. Probe-verified catalog
connectors (and demo apps) may honor the upstream’s own
readOnlyHint. Custom URL apps are stricter: an upstream hint
counts only when the tool name also looks read-only (prefixes such as get_, list_, search_). Anything not
provably read-only counts as a write and is withheld under a read-only pass. The name heuristic is best-effort, not a
guarantee that a tool is safe.Per-tool control and drift
Within an app you can turn individual tools on or off, workspace-wide. On Enterprise, you can also choose whether newly discovered tools default on or off. Passport snapshots each integration’s tool set and detects drift, so you are alerted when an upstream MCP or reviewed API manifest changes rather than silently gaining new capabilities. With Keep off until I approve selected, a new tool stays off until you approve it. A changed tool stays off too, unless the change is low risk: the tool reads data before and after, its new description and parameters pass Passport’s poisoning and runtime security checks, and it only gains optional parameters or a reworded description. Low-risk changes keep serving and are still recorded in the tool history and alert as auto-approved. Tool history and in-app alerts are part of Pro and Enterprise. A write tool, or a change that removes, retypes, constrains, or requires a parameter, stays off. On every plan, a tool whose new definition looks like a poisoning attempt is switched off until you review it. Held tools are listed under Review tool changes in Apps → your app → Connections, with what changed. Select the ones to approve and turn them back on in one step. Flagged tools are approved one at a time. Until then, an agent that calls a held tool is told the app changed it and an admin review is pending, and Activity records the call as held for admin review rather than as a tool you switched off.Action approvals
Separate from catalog requests, a member’s own Safety preference can put a write or destructive action on hold until that person allows it. Passport asks them wherever it can reach them: agents that support MCP elicitation get a prompt inside the agent itself, with the app, the action and its risk class, the account, and a redacted one-line argument summary; everything else waits while the person decides in the Passport app. Either way the decision lands on the same durable approval record, so a decision made on one replica releases a call held on another. The audit trail records how consent was collected. An approval shows whether it was answered in Passport or in an agent, and which agent install answered it; the Activity row for the released call carries the same provenance next to the approval it consumed. Autonomous agent credentials are never prompted, and the CLI keeps returning an explicit pending result for its own retry loop. SetPASSPORT_APPROVAL_ELICITATION=off to turn in-agent prompts off for a deployment and keep the Passport-app flow as the only channel.