klypix-mcp

An open, local-first, agent-neutral canvas file your AI reads and writes over MCP — works with Claude, Cursor, Cline, any model.

Your AI forgets everything between sessions, and you can't hand it your whole messy project at once. klypix-mcp fixes that with a single portable file:

• .klypix is one file that holds your whole project spatially — text, PDFs, screenshots, audio, code, links — arranged and connected on an infinite canvas.

• This package is an MCP server: any MCP-capable agent (Claude Desktop, Cursor, Cline, …) can list, read, search, create, and append to your .klypix canvases — so your agent gets durable, multimodal, spatial memory.

• Local-first + agent-neutral. The file lives on your disk. No lab is in the loop — read it with Claude today, GPT tomorrow, a local model next week. No vendor can take it away.

Quick start (60 seconds)

# point it at any folder of .klypix files (a "vault")
npx klypix-mcp --vault ./canvases

Then add it to your MCP client. For Claude Desktop, in claude_desktop_config.json:

{
  "mcpServers": {
    "klypix": {
      "command": "npx",
      "args": ["-y", "klypix-mcp", "--vault", "/absolute/path/to/canvases"]
    }
  }
}

Now ask your agent things like "summarize the canvas roadmap," "turn these notes into a board," or "add a card with the decision we just made."

Tools the server exposes

Use it as a library

import { parseKlypix, buildKlypix, appendToKlypix, structToMarkdown } from 'klypix-mcp';

…or read/write from the shell:

npx -p klypix-mcp klypix-read   path/to/board.klypix      # → markdown brief
echo '{ "title": "Plan", "cards": [{ "text": "kickoff" }] }' \
  | npx -p klypix-mcp klypix-write --out plan.klypix

The .klypix format

A .klypix is just a ZIP of JSON + assets — open, inspectable, versioned. Full spec in FORMAT.md. The point: you own the file, and any agent can drive it.

Why this exists

Frontier labs are racing to put your context inside their canvas — a roach motel your work checks into and never leaves for a competitor. klypix-mcp is the opposite: your project, your file, any model, offline. That's the one thing a lab is structurally disincentivized to build.

MIT licensed. Built by Dahshan Labs. The KLYPIX desktop app is the spatial editor for these files — but the file and this server are fully open and work without it.