Personal
FigJam Context MCP: Boards as LLM Context
An open-source MCP server, published on npm and in the official MCP Registry, that turns busy FigJam boards into compact context an LLM can search and cite.
After each university course we exported our FigJam boards as PDFs to write documentation with AI, which wasted tokens and lost structure. I had the idea in autumn 2025 and built the server from July 2026, partly to learn MCP.

The short version
- Five MCP tools: ingest, read, ask with sources, compare versions, check the setup
- Every answer cites node IDs with links back to the board
- After an edit, diff_board names the changed sticky or table cell
What I worked on
- Geometry groups the board, arrows become relationships, and vision runs only on images.
- I ran it with AI agents: Claude planned and reviewed, coding agents implemented, and I decided and merged.
What came out of it
I shipped 19 releases in October and November 2026, published by CI to npm and the MCP Registry, with about 1,900 tests. The demo below is a real run of 0.14.0, which completes the project.

What I learned
In November I benchmarked my server against Figma’s MCP server, which had added FigJam reading in 2026. An agent documented a real board with both. Where code could run, Figma’s server returned the whole board and the agent read it all. Mine shortened much of it. Both found the same contradictions. Figma covers this well, so I closed the project. In a field this fast, I now benchmark against the platform’s tools throughout a project.
Tools and methods
- TypeScript
- Model Context Protocol
- Figma REST API
- Vision LLMs
- Zod
- Vitest