settings / mcp-server

Model-Context-Protocol-Endpunkt für KI-Agenten

Master switch for the MCP endpoint. When off, /mcp returns 404 and discovery reports disabled: true. Defaults on for Pro+ editions.

Allow anonymous (no API key) callers to discover and use tools whose collection has mcp.access: 'public' set on its schema. Turn this off to require an API key for every MCP request, regardless of collection settings.

Origins permitted to call /mcp from inside a browser tab. Pick a known browser-based AI client from the dropdown, or type a custom origin (e.g. https://example.com) for in-house playgrounds.

Native MCP clients (Claude Desktop, Cursor, Claude Code, MCP Inspector) connect regardless of this setting — CORS is a browser-only mechanism and never reaches non-browser clients. Server-side AI integrations are also unaffected.

Leave empty (default) to block browser clients entirely. Use * to allow any origin (not recommended for sites with non-public collections — any website's JavaScript could read your public MCP surface).

Optional prefix prepended to every MCP tool name. Useful when running multiple T3 sites in the same AI agent so tools are distinguishable (e.g. bistro produces bistro_list_collections, bistro_query_collection). Leave blank for default names. Snake-case only, max 24 characters. Note: changing this requires connected AI clients (Claude Desktop, Inspector, etc.) to reconnect — existing sessions will see stale tool names until they re-initialize.

When on, AI agents can call resources/subscribe on URIs like tcms://blog/ and T3 will push notifications/resources/updated when matching content changes. When off, the SDK still accepts subscribe calls but T3 won't push notifications — the kill switch for production troubleshooting.