MCP

MCP

Model Context Protocol — agents query your synthesized knowledge

Most knowledge tools hand an AI a giant blob of raw text and hope it makes sense of it. MCP (Model Context Protocol) gives agents a structured catalogue of your knowledge — search, list_sources, get_entity, traverse_graph, summarize. Claude, Cursor, and other MCP clients call those tools the same way they'd call an API, but with your permission, scoped tokens, and an audit trail. The result: AI assistants that read what's already organized, not what they have to organize from scratch.

Key capabilities

  • Remote OAuth connector — https://mcp.deep-thought.cloud/mcp. Full OAuth 2.1 dynamic client registration: browser-based consent → token. No CLI needed; one click in Claude Desktop.
  • Local stdio MCP — for headless agents and developer environments. deep-thought-mcp pair produces a scoped device token.
  • 17 synthesized tools — search, list_sources, get_source, get_entity, traverse_graph, apply_tag, create_action, add_note, summarize, and more. Agents query intent, not raw blobs.
  • Scoped device tokens — kb:read, kb:write, graph:read. Revoke per-device in Settings → Connected devices on app.deep-thought.cloud.
  • Per-tag context boundaries — chats stay scoped to the tags you authorize; the agent can't peek beyond.
  • Audit trail — every tool call logged with req_id, latency, and source. The same instrumentation that powers /mcp-stats in the admin dashboard.

Specialized applications

  • Claude Desktop — chat with your knowledge graph, no CLI required
  • Claude Code — terminal-native AI development, with your vault one tool-call away
  • Cursor — your knowledge inline while coding
  • Custom MCP clients — anything that speaks the protocol
Access MCP → View all services