MCP server
MDDock ships a built-in Model Context Protocol server. Point Claude Desktop, Cursor, or any MCP-aware client at it and your agent gets four unified tools backed by the local vault.
Start the server
Wire it into Claude Desktop
Add this to
(macOS) or the equivalent on your platform:
Restart Claude. The four tools below should appear in the MCP picker.
The four unified tools
mddock_recallβ Search the workspace for relevant chunks. Supports hybrid (default), lexical-only, vector-only, graph expansion, and entity extraction. Also implements mini-KnowQL via optional shape/confidence_min/budget_tokens.mddock_readβ Read a markdown surface by path, md:// URI, audit events, harness package, or installed memory packs.mddock_rememberβ Persist a workspace-mutating fact: write a memory page, track an agent actor, or edit an existing doc with audit logging.mddock_proposeβ Propose a workspace change: marketplace search/inspect, install a module, resolve a sync conflict, or suggest a wikilink between docs.
Read-only by default
Mutating actions go through `mddock_remember` (direct edits with audit logging) or `mddock_propose` (marketplace installs, conflict resolution, and suggested links). Proposals are surfaced in the desktop client for review before they land, so an agent can't silently rewrite your notes.
Cursor / Continue / others
Any MCP-compliant client works. The transport is plain stdio JSON-RPC; no MDDock-specific glue required. See /docs/agent-protocol for the read-only HTTP surface (no MCP needed).
Troubleshooting
- β’mddock doctor, checks workspace, schema, and binary versions
- β’Logs land at ~/.mddock/logs/mcp.log
- β’If a tool is missing in the client, try mddock mcp --json from a terminal to confirm it boots