Connect External Chats

Connect Claude Code and other MCP-capable tools to your RA-H graph

What This Does

This lets external assistants read and write to your RA-H graph through MCP.

Typical uses:

  • search your existing graph
  • ground broader work in prior graph context
  • add or update durable nodes
  • propose and confirm edges

The standalone MCP server writes node data directly to the local SQLite database. It does not create chunk rows or embeddings by itself. The app-owned pipeline later turns nodes.source into readable chunks, full-text indexes, node-level vectors in vec_nodes, and passage vectors in vec_chunks.

Use the standalone MCP server with a pinned version:

{
  "mcpServers": {
    "ra-h": {
      "command": "npx",
      "args": ["--yes", "ra-h-mcp-server@2.1.2"]
    }
  }
}

After editing config:

  • fully restart the client
  • run RA-H once first so the DB exists
  • verify the tools are visible

Expected Agent Behavior

  • use queryNodes first for specific-node lookup
  • use retrieveQueryContext for broader graph grounding
  • use getContext only for orientation
  • search before creating
  • propose durable writeback selectively
  • propose likely edges before creating them
  • do not assume a newly-created MCP node is immediately chunked or vectorized
  • do not treat chunk_status = chunked as proof that every historical chunk has a row in vec_chunks

Optional Memory Reinforcement

You can add one short reinforcement line in your client memory file, but the MCP contract should still work without prompt surgery.

Recommended pattern:

Retrieve relevant RA-H context before substantive work, search before creating, and keep durable writeback prompts brief and confirmation-gated.

More Detailed Open-Source / Local Guidance