Sources: APIs & integrations

The station's tools: the tenant's APIs, and other stations called as APIs — the integrations. Each operation says whether it reads or writes, its risk and what comes out of it.

The APIs

apis:
  - name: stripe
    openapi_url: https://payments.example.fr/openapi.json
    base_url: https://payments.example.fr
    auth: {type: api_key, key_name: Authorization, key_location: header, key_env: STRIPE_KEY}
    operations:
      - {id: getInvoice,     access: read}
      - {id: createRefund,   access: write, risk: high, reversible: false}
      - {id: createCreditNote, access: write, risk: medium, reversible: voidCreditNote}
      - {id: voidCreditNote, access: write, risk: medium}
      - {id: deleteCustomer, access: write, risk: critical, exposed: false}
  - name: zendesk
    auth: {type: oauth2, on_behalf_of: identity}   # the tool sees WHO acts
    operations:
      - {id: sendReply, access: write, risk: medium, audience: external}

attribute	effect
access	read or write. A read runs right away; a write is reserved.
risk	low · medium · high · critical. Weighs on warnings (an agent allowed to submit a high write).
exposed: false	listed so it is documented, but never reachable.
audience: external	the write leaves the organisation (an e-mail to the customer). It cannot carry confidential data.
reversible	the operation that undoes it, or false. The approval screen says "irreversible"; at submit, irreversible writes go last.
auth.on_behalf_of: identity	the call carries the person's identity, not a shared key.

A station as an API

apis:
  - name: fournisseur-materiel
    station: mcp.convoy.sale/materiel-pro/mcp

The call goes through the other station: its rules apply. Our agent becomes a sub-identity there. The two logs share a correlation identifier.