Connect AI apps (MCP)

​

Server URL

https://api.zeeg.me/mcp

​

Connect from your AI app

claude mcp add --transport http zeeg https://api.zeeg.me/mcp

{
  "mcpServers": {
    "zeeg": {
      "url": "https://api.zeeg.me/mcp"
    }
  }
}

{
  "servers": {
    "zeeg": {
      "type": "http",
      "url": "https://api.zeeg.me/mcp"
    }
  }
}

​

Sign-in and consent

​

Plan requirement

• find_available_times, diagnose_availability and book_meeting need a paid plan (not trial) — the same premium-scope check the public API applies to available-time and booking-creation requests.
• invite_users and revoke_invitation need a paid plan (not trial), as the users:write scope does on the public API: an accepted invitation adds a paid seat.
• The voice agent tools (list_voice_agents, get_voice_agent, list_voice_agent_options, create_voice_agent, update_voice_agent, add_agent_route, update_agent_route, delete_agent_route, list_agent_calls, get_agent_call and start_outbound_call) need AI minutes in the workspace: an AI minutes subscription or enough AI credit.

​

Tools

​

Safety

• Confirmation before it matters. book_meeting, cancel_booking, reschedule_booking, hand_over_booking, and start_outbound_call are marked destructive, so a well-behaved AI app confirms with you before calling them — booking, cancelling, moving, handing over and calling all have a real-world effect, and start_outbound_call spends AI minutes. update_schedule, add_time_off, upsert_crm_records, update_note, delete_note, set_event_type_active, update_event_type, replace_event_type_questions, update_event_type_notifications, update_routing_form, replace_routing_form_questions, replace_routing_form_routes, update_team, remove_team_member, revoke_invitation, update_voice_agent, update_agent_route and delete_agent_route are marked destructive too, because they overwrite or remove what’s already there, take a booking page or form offline, break a shared link or change what a voice agent says on live calls. So is set_invitee_no_show, because a no-show can start workflows that email the invitee, and so is invite_users, because an invitation email cannot be called back. add_note, create_single_use_link, create_event_type, duplicate_event_type, create_routing_form, duplicate_routing_form, create_team, add_team_member, create_voice_agent and add_agent_route change data without being marked destructive, since they only add; their descriptions still ask the AI app to confirm with you, because a new scheduling page, form, voice agent or route goes live. delete_event_type, delete_routing_form and delete_team are marked destructive and say in their description what is lost for good: a deleted routing form takes every submission with it, and a deleted scheduling page takes the voice agent routes that book it. Each asks the AI app to confirm with you by name first, and to turn a page or form off instead when you only want it offline.
• Text from outside your workspace is marked as data, not instructions. Booking answers, notes, and routing form submissions are written by invitees and form respondents. The server labels this content and tells the connected AI app never to treat it as instructions — a booking answer that says “ignore your previous instructions” stays inert.
• Every tool call is logged. Which app, which tool, what it touched, and whether it succeeded — visible to you (and, for admins, the whole workspace) at Settings → Connected apps. Arguments and results aren’t stored, only what happened.
• Lean output by default. Tool results return what the task needs — a list of bookings, for instance, doesn’t include phone numbers or form answers unless the app explicitly asks for them.

​

Data processing

​

For OAuth client builders

• Discovery: GET /.well-known/oauth-authorization-server and GET /.well-known/oauth-protected-resource/mcp, both under https://api.zeeg.me.
• Resource: the protected-resource metadata’s resource is exactly https://api.zeeg.me/mcp.
• PKCE: required, S256 only. An authorize request with plain, or without code_challenge_method, is refused with invalid_request.
• Issuer identification (RFC 9207): every authorization response carries iss, equal to the metadata issuer (authorization_response_iss_parameter_supported: true).
• Dynamic client registration: POST /oauth/register, open to any redirect host — the consent page names the redirect host and marks the app unverified unless it’s on our first-party allowlist, so don’t rely on registration itself as a trust signal.
• Client ID metadata documents (CIMD): also supported (client_id_metadata_document_supported: true), and preferred over DCR for a high-traffic client — the metadata is fetched from the client’s own client_id URL instead of registering a new row per connection.
• Client type: public clients only — token_endpoint_auth_methods_supported: ["none"]. There’s no client-secret flow.
• Scope: every MCP client is pinned to mcp:use regardless of what it requests. A token that carries mcp:use is rejected by every /v1 and /v2 REST route, and a normal API token is rejected by the MCP endpoint — the two token types don’t cross over.
• Token lifetimes: access tokens expire after 1 hour; refresh tokens after 30 days and rotate on use — reusing a spent refresh token returns invalid_grant.
• Loopback redirects: http://localhost/callback and http://127.0.0.1/callback match on any port, for local tools like Claude Code.
• Token endpoint: accepts application/x-www-form-urlencoded, per the OAuth spec.
• Tool results: every tool except get_zeeg_guide returns structuredContent described by its outputSchema in tools/list, plus the same JSON as a text block for clients that read only text.