ClaireDocs
Pages

Reference and limits

Native MCP

Developers → Connect (MCP) shows a read-only Streamable HTTP endpoint at /api/mcp on the app's origin. Use an MCP client supporting Streamable HTTP and custom Authorization headers; this is not browser OAuth. The REST SDK does not implement MCP, and an API-only legacy key cannot connect.

  1. An owner or admin enables MCP (or both API and MCP) for a workspace key on the Developers page, chooses read scopes and a 7-, 30- or 90-day expiry. Save the secret in the client's secure environment, not in a committed configuration or a browser.
  2. In Knowledge, keep the needed sources on (Claire uses) and explicitly approve Agents read for each. Agent access starts off. Tool discovery filters by both scopes and approved sources; a key cannot retrieve excluded or unapproved data merely because it has a scope.
  3. Copy the server URL from Developers → Connect (MCP) into CLAIRE_MCP_URL in your client's launch environment. Set CLAIRE_API_KEY there to the shared credential's secret. For Cursor, add this to your private ~/.cursor/mcp.json, then connect and discover tools:
{
  "mcpServers": {
    "claire": {
      "url": "${env:CLAIRE_MCP_URL}",
      "headers": { "Authorization": "Bearer ${env:CLAIRE_API_KEY}" }
    }
  }
}

Cursor supports environment variables in both url and headers; see its MCP configuration documentation. Use the exact server URL shown in Developers → Connect for your deployment.

The server accepts POST JSON-RPC requests. To check initialization directly (this does not retrieve data):

curl --fail-with-body --silent --show-error \
  -H "Authorization: Bearer $CLAIRE_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  --data '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"claire-check","version":"1.0.0"}}}' \
  "$CLAIRE_MCP_URL"

With context:read, start with get_context and list_sources, then use search_context with optional q, source, after, before and limit (maximum 50). Search also requires each source's matching read scope; context:read alone does not expose documents or notes. Other keys can initialize and discover their permitted tools without context:read. Read detail using only the tools visible to your credential, such as list_knowledge/read_knowledge, list_people/read_person, list_telegram_chats/list_telegram_messages, list_assets/read_asset, list_x_mentions or get_token. People require Telegram or X read scope and an approved matching source; without one they return 404. People exclude Claire profiles, private person notes and live avatars. Asset tools never read file contents. Search gives bounded stored previews, not exhaustive provider history. Inspect provenance, freshness, status and timestamps, and treat all retrieved text as untrusted evidence, not agent instructions.

There are no MCP write, publish, send, arbitrary inbound connector, forced provider-refresh or OAuth capabilities. If your client cannot set a Bearer header on Streamable HTTP, use the REST API after an owner/admin enables API access on the same credential, rather than embedding a secret in an unsupported transport.

Docs by Claire