Connect an AI agent to OwlCAD

Point Claude Code, Claude Desktop or Cursor at OwlCAD so it can build and check models for you.

OwlCAD speaks MCP, the protocol Claude Code, Claude Desktop and Cursor use to reach outside tools, so an agent can build print-ready models without a browser. It is a Pro feature and spends no credits: the agent you are already talking to authors the model itself, and OwlCAD does the geometry. The tools cover the whole loop — evaluate runs the real CAD kernel and reports whether the shape is solid, how many triangles it has and how big it is; render returns pictures the agent can actually look at, which is the only way it can tell a shape that computes from a shape that is right; print_check reports overhangs, thin walls and whether it fits your printer; export returns an STL, 3MF or OBJ. Saving gives you an ordinary OwlCAD project — it appears in your dashboard with a preview, and you can open and edit it like anything else. Everything is millimetres, Z-up, with rotations in degrees. The three clients are not wired up the same way: Claude Code and Cursor accept the address of a remote server directly, but Claude Desktop’s config file only knows how to launch a program on your own machine, so it runs a small bridge called mcp-remote that carries the connection out to us. That is why its block looks nothing like the other one, and why pasting the short version into it fails without ever saying why. If you ask for changes over several turns, tell the agent to use update_project rather than save_project, so you keep one project and one link instead of a new one each round. There is a second way in, for an agent working in a BROWSER rather than headless — Claude in Chrome, or any computer-use loop driving a real page. Unlike the MCP route it is free and needs no token — it runs in your own browser. Open the editor at /app?agent=1 and it gains a scripting bridge on the page, so the agent calls the editor directly instead of guessing pixel coordinates on a 3D canvas that has no accessibility tree, and it can take a screenshot to see what it actually did. Reading is always available, so an agent can inspect any editor session; the parameter is what lets it EDIT, it applies to that tab only, and /app?agent=0 turns editing off again. Everything it does is an ordinary undoable edit, so Ctrl+Z takes back anything you did not want.

1. Create an access token with the scopes the agent needs — read alone is enough to look at your work.
2. Claude Code and Cursor reach the server directly: the URL is https://mcp.owlcad.com, and the token goes in an Authorization header as “Bearer …”.
3. Claude Desktop cannot — its config file only launches servers on your own machine — so it runs the mcp-remote bridge instead: the command is npx, the arguments are -y mcp-remote https://mcp.owlcad.com --header Authorization:${AUTH_HEADER}, and the token goes in env as “Bearer …”. Copy the exact block from owlcad.com/mcp-server: the missing space after the colon is deliberate, because Claude Desktop on Windows mangles a header argument that contains one.
4. Restart the client. Ask it to read owlcad://schema/primitives first — that lists every shape and the valid range of every setting.
5. Ask for a model. A good loop is: build it, run evaluate to check it is solid, render it to look at it, then save_project.
6. Ask for the editor link when you want to take over — open_in_editor hands back a URL that opens the model in OwlCAD.