compy-mcp

Let your AI agent search and read the notes you captured on the web with Compy.

Compy is a local-first Chrome extension for highlighting and note-taking on any web page. compy-mcp is a Model Context Protocol server that exposes those notes to an MCP client (Claude Code, Claude Desktop, …) so your agent can pull the relevant highlights and tasks on demand — no copy-paste.

100% local. Read-only. No account, no network — it just reads a JSON file that Compy writes on your machine.

How it works

Compy (browser)  ──"Sync for AI"──▶  ~/Downloads/compy/compy-export.json
                                              │  (watched)
                                       compy-mcp (stdio)
                                              │  tools + resources
                                        Claude Code / any MCP client

1. In Compy's dashboard, open the Copy for AI menu → Sync for AI. This writes ~/Downloads/compy/compy-export.json. Re-run it whenever you want the agent to see your latest notes.
2. Register this server with your MCP client (below).
3. Ask your agent things like "check Compy for anything I flagged on the billing page" — it calls search_notes and gets your notes, quotes, tags and URLs.

Install

Requires Node 18+.

Claude Code

claude mcp add compy -- npx -y compy-mcp

Claude Desktop / other clients

Add to your MCP config:

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

Custom file location

By default it reads ~/Downloads/compy/compy-export.json. Override with an env var or flag:

COMPY_FILE=/path/to/compy-export.json npx compy-mcp
# or
npx compy-mcp --file /path/to/compy-export.json

Tools

Tool	What it does
search_notes	Ranked keyword search over note text, quotes, titles and tags. Filter by site / label.
list_tasks	List notes/tasks, optionally filtered by site / label.
list_sites	Sites you have notes on, with counts. Good first call.
get_note	Fetch one note by id.
get_markdown	Render notes as a Markdown task list (optionally scoped to a site).

Resource: compy://sites — overview of sites + counts.

Why keyword search, not embeddings?

The note set is small (tens to a few hundred short notes) and the calling model is the semantic layer — it reasons over what search_notes returns. Fast substring/token ranking is plenty at this scale and keeps everything local, with zero setup and no API keys.

License

MIT