STACK runs between AI agents and the systems they use. Each agent works under a short-lived Passport that states what it may do. STACK checks every call against that Passport, records suspicious behavior, can deny configured actions and writes each decision to a hash-chained audit log.
This connector brings STACK's governance into the conversation. With it you can:
- register an agent, and list your agents, their permissions and any that are blocked
- grant or narrow an agent's access to a connected service. A new grant, or any widening, needs a human to approve it in the STACK Console; narrowing applies immediately
- see active Passports and revoke one, every Passport for an agent, or all of them
- open a Mission with a spend or action limit, complete or revoke it, and see what an agent did in it
- decide flagged Passport checkouts waiting for review
- read the audit log, verify its hash chain and export signed records
- list security events, adjust detector and PII-redaction settings (loosening needs human approval) and scan text for prompt injection
- manage notification rules, inbound webhooks, drop-offs and fleets
- read your white-label partner tenants and Connect grants
- post and browse public skill requests
- check usage
A revocation applies to the agent's next call through STACK and cascades to any Passports it delegated.
Two tools reach outside STACK: verifying a connection makes one test call to the provider with the stored credential (never shown), and scanning for prompt injection sends flagged content to an LLM classifier on OpenRouter (Claude Haiku 4.5); STACK keeps the verdict and matched snippets, not the content. The connector does not connect or display credentials, issue Passports or run paid skills; those stay in the STACK Console, CLI and full MCP endpoint. Sign-in uses OAuth; a token issued to this connector is refused by the full endpoint.
Requires a STACK account. Setup: https://getstack.run/docs/mcp-setup#directory
