origin-mcp

MCP server for Origin. Persistent memory across Claude, ChatGPT, and Cursor.

Origin is a local-first companion for people who work with AI every day. Conversations across tools become connected, deduplicated, and editable. origin-mcp is the bridge: it lets any MCP-compatible tool read and write to your shared memory through the Model Context Protocol.

Install

Add to your MCP config (Claude Code, Cursor, Claude Desktop, Windsurf, Gemini CLI):

{
  "mcpServers": {
    "origin": {
      "command": "npx",
      "args": ["-y", "origin-mcp"]
    }
  }
}

Or install the binary directly:

# Via Homebrew
brew tap 7xuanlu/tap
brew install origin-mcp

# Via cargo
cargo install origin-mcp

Then add the binary path to your MCP config:

{
  "mcpServers": {
    "origin": {
      "command": "origin-mcp"
    }
  }
}

How it works

origin-mcp connects to the Origin daemon running on 127.0.0.1:7878. The daemon owns all storage, embeddings, and refinement. This server is a thin MCP interface to it.

Claude Code / Cursor / Claude Desktop
    |
    | MCP (stdio)
    v
origin-mcp
    |
    | HTTP
    v
Origin daemon (origin-server)
    |
    v
Local SQLite + embeddings + knowledge graph

If the daemon isn't running, npx origin-mcp starts it automatically.

Tools

Tool	What it does	Annotations
`remember`	Store a memory, fact, preference, or decision. The backend auto-classifies type, extracts entities, and links to the knowledge graph.	write, non-destructive
`recall`	Search memories and knowledge graph by natural language. Returns ranked results with source tracing.	read-only
`context`	Load session context: identity, preferences, goals, and topic-relevant memories. Call this at session start.	read-only
`forget`	Delete a specific memory and clean up entity links. Requires the memory ID.	destructive, idempotent

What agents should know

The server ships with proactive-capture instructions that guide agents to store the right things at the right granularity. Key ideas:

Two mental models: profile (about the user) vs knowledge (about the world). Agents should think in these terms when deciding what to store.
One idea per memory. "Prefers TDD" and "uses pytest" are two memories, not one. Specific memories retrieve better than broad summaries.
Include the why. "Switched to dark mode because of migraines" is more useful than "uses dark mode."
Omit memory_type. Let the backend auto-classify. Agents get it wrong more often than the classifier.
Anti-noise rules. Don't store conversation filler, tool output, or things trivially re-derivable from code.

See src/tools.rs for the full with_instructions text that agents receive.

Options

--origin-url <URL>    Override Origin server URL (default: http://127.0.0.1:7878)

What Origin does with your memories

Origin doesn't just store what agents send. A background engine refines memories over time:

Deduplication. Overlapping memories are merged automatically.
Concept distillation. Related memories are clustered into concepts: compact, wiki-style summaries that save tokens on retrieval.
Knowledge graph. Entities and relations are extracted and linked, so "Alice leads the deploy refactor" connects Alice, the project, and the decision.
Contradiction detection. When new information conflicts with existing memories, Origin surfaces it for your review.

The longer you use it, the better the retrieval gets.

Requirements

Origin daemon running locally (via the desktop app or origin-server install)
macOS Apple Silicon (M1+) at v0.1.0. Linux x64 binaries are built but not yet tested in production.

License

MIT

origin-mcp 0.1.2