If you use Claude or ChatGPT, use the consumer flow instead —
Connect Claude. This is for your own clients.
Option 1 — an MCP-aware client
Clients that already speak MCP (e.g. desktop AI apps with custom connectors) only need a connector’s URL from the portal — each connector is its own endpoint, and/mcp is the Concierge. They discover the authorization
server themselves, take you through login, and store your token.
Option 2 — mcp-remote (stdio → remote bridge)
Clients that can only run local stdio servers can bridge to Consile withmcp-remote. It opens the browser for OAuth the first time and caches your token:
Option 3 — programmatically via an MCP SDK
Use an official MCP SDK with a Streamable HTTP transport and an OAuth provider. Once connected, the flow is the standard MCP handshake:tools/call
Call a tool with its arguments — see the
API reference.
Example: raw JSON-RPC
Atools/call request over POST /gsc/mcp (abbreviated — the
Authorization: Bearer <token> header is omitted here). Note top_queries is a
GSC tool, so it goes to the GSC endpoint, not the aggregate /mcp:
If you call a connector endpoint without a valid token, you get
401 with a
WWW-Authenticate header pointing to the discovery document. That’s the normal
starting point for OAuth — see Authentication.