generated: '2026-08-29'
method: probed
source: >-
  https://mcp.relevanceai.com/.well-known/oauth-authorization-server,
  https://mcp.relevanceai.com/.well-known/oauth-protected-resource,
  https://relevanceai.com/docs/integrations/mcp/mcp-server
provider: Relevance AI
providerId: relevance-ai
description: >-
  The REST API declares no oauth2 security scheme — it authenticates with a static
  `Authorization: project:api_key` header and has no scope concept at all. The only OAuth
  surface is the MCP server, and its scope model is deliberately coarse: one scope, with real
  authorization carried by the caller's project role rather than by scope strings.
rest_api:
  oauth2: false
  scopes: []
  detail: >-
    derive-oauth-scopes.py over openapi/relevance-ai-openapi.json found 0 oauth2 flows across
    566 operations. The single securityScheme is apiKey in header. An API key grants
    unrestricted access to every resource in its project — the docs say so explicitly — so
    there is no way to issue a least-privilege REST credential.
mcp:
  authorization_server: https://mcp.relevanceai.com/
  scopes_supported:
    - name: claudeai
      description: >-
        The only scope the authorization server advertises, in both the authorization-server
        and protected-resource metadata. It carries no capability semantics.
  grant_types: [authorization_code, refresh_token]
  pkce: S256
  dynamic_client_registration: true
  effective_authorization:
    model: project role, enforced server-side
    detail: >-
      Because there is one scope, the real permission boundary is the caller's Relevance AI
      project role, applied at consent time. Write and destructive tools are REMOVED from the
      tool list server-side for restricted sessions rather than merely hidden client-side.
    roles:
      - role: Admin
        access: full
        can_opt_into_run_only: true
      - role: Editor
        access: full
        can_opt_into_run_only: true
      - role: Member
        access: full
        can_opt_into_run_only: true
      - role: Viewer
        access: run-only
        can_opt_into_run_only: forced
      - role: Chat
        access: chat-only
        can_opt_into_run_only: forced
    consent_toggle: >-
      The OAuth consent screen shows a Run-Only mode toggle for Member, Editor and Admin,
      defaulting to off (full access).
    limitation: >-
      Chat-role users cannot set up dynamic (user-level) OAuth connections via MCP.
  session_scope: one Relevance AI project per authenticated connection
docs: https://relevanceai.com/docs/integrations/mcp/mcp-server#access-control
gaps:
  - REST API keys cannot be scoped or least-privileged; they are all-or-nothing per project.
  - A single OAuth scope means a client cannot request narrower access than its role allows.
  - No scope reference page is published; the scope name was read from the metadata document.