Node-based parametric CAD on build123d. Wire nodes, get solid geometry — transpiled to Python, rendered live, and drivable by an AI coding agent.
Sending a drawing to an agent needs noodle running on your machine. The quickest way: copy this and give it to your coding agent (Claude Code, Codex, Cursor…). It reads the repo first, stops if Docker is missing, installs noodle and connects itself — then "look at what I marked" works.
Install noodle (https://github.com/rederyk/noodle) on this machine and connect yourself to it.
1. Check it first: read README.md ("Install & run"), AGENTS.md and the security note — noodle runs graph code as unsandboxed Python, so it is for local, single-user use only. Tell me if anything there worries you before going on.
2. Check that Docker is installed and running (`docker info`). If it is not, tell me how to install it on my OS and stop there.
3. Clone the repo and start it: `docker compose up -d --build` (the first build downloads ~1 GB and takes a few minutes). Wait until http://localhost:8090/health answers.
4. Connect yourself through its MCP server — for Claude Code: `claude mcp add noodle -- docker exec -i noodle python mcp_server.py`; for other agents use the config in AGENTS.md. Then call `cad_help` and read it.
5. Tell me the editor URL (http://localhost:8090/nodes) and that, from now on, I can draw on a part in the viewer (✎ Disegna) and ask you "look at what I marked" — you read it with `cad_notes`.
or install by hand
An agent calls cad_snapshot and sends a URL. The link stays on that result while
the workflow moves on.
Agents write long CodeBlocks. Instead of making them assemble catalogue nodes, noodle reads the block's structure from its AST — the code is never rewritten.
cad_codeblock_sections.An agent edits the graph over MCP or HTTP while you have it open.
/measure and /lint for geometry facts.