Skip to content

A desktop coding agent with free access to frontier models.

TypeScript
3.0 KB
Raw Blame

Den API MCP Exposure Policy

The MCP catalog is generated from openapi.json, then filtered by policy.ts before tools are registered.

For third-party client setup and invalid_target troubleshooting, see Connect a third-party MCP client with OAuth.

Allowed Tags

Every tagged Den API product surface is allowed unless it is listed under blocked tags or blocked operation IDs:

  • API Keys
  • Capability Sources
  • Config Objects
  • Connectors
  • Desktop Policies
  • GitHub
  • Invitations
  • LLM Providers
  • Marketplaces
  • Members
  • Organizations
  • Plugins
  • Roles
  • Teams
  • Users
  • Worker Activity
  • Worker Runtime
  • Workers

Desktop Policies reads require org admin; mutations require super-admin + Enterprise entitlement. Both are enforced in-route.

Blocked Tags

These tags are intentionally excluded from MCP:

  • Admin: internal administrative controls should not be broadly exposed as agent tools.
  • Authentication: OAuth/session plumbing is used to authorize MCP, not exposed through MCP.
  • System: health, docs, and other service metadata are not product actions.
  • Webhooks: external webhook ingress routes require provider signatures and should not be invoked by agents.

Blocked Paths

Routes are blocked if their path:

  • starts with /api/auth
  • contains /admin
  • contains /webhooks

This catches auth/admin/webhook routes even if they are untagged or incorrectly tagged.

Blocked Operation IDs

These individual operations are blocked even though their tags may otherwise be allowed:

  • postApiKeys: creating API keys returns credentials and should stay behind explicit UI/API flows.
  • postV1ApiKeys: generated OpenAPI ID for creating API keys.
  • deleteApiKeysByApiKeyId: destructive credential revocation should stay behind explicit UI/API flows.
  • deleteV1ApiKeysByApiKeyId: generated OpenAPI ID for deleting API keys, if present.
  • deleteOrg: defensive block for organization deletion if a route is added.
  • deleteV1Org: defensive block for organization deletion if a route is added.
  • deleteV1OrgsByOrgId: defensive block for organization deletion if a route is added.
  • postWorkersByWorkerIdTokens: worker token minting returns credentials and should stay behind explicit UI/API flows.
  • postV1WorkersByWorkerIdTokens: generated OpenAPI ID for worker token minting.
  • postOauthProvidersByProviderIdDisconnect: removing a connected OAuth credential is a mutation that should stay behind explicit UI/API flows, not an agent-callable tool.
  • postV1OauthProvidersByProviderIdDisconnect: generated OpenAPI ID for disconnecting an OAuth provider.

Untagged Operations

Untagged operations are excluded by default. Today these are OAuth/MCP discovery and registration routes, for example:

  • /.well-known/oauth-authorization-server
  • /.well-known/openid-configuration
  • /.well-known/oauth-protected-resource
  • /api/auth/oauth2/authorize
  • /api/auth/oauth2/register
  • /register

They are required for OAuth/MCP setup, but should not appear as callable MCP tools.