ParCAD MCP server

ParCAD

Parametric CAD you write as code — for people, and for agents.

Status: 0.0.7, experimental. One author, and the DSL still moves. Every operation is checked against measured geometry (eval/cases/), but none of it has had the years of abuse that make a CAD kernel trustworthy. Measure a part before you machine it.

const plate = box(80, 60, 8).tag("plate");
const wall = box(8, 60, 40).at(-36, 0, 20).tag("wall");

const body = union(plate, wall, { blend: 6 });          // the blend is the fillet
const drilled = body.cut(...grid(2, 2, 36, 40).map(([x, y]) => cylinder(3, 32).at(x, y)));

return drilled
  .edges({ curve: "circle", role: "hole", adjacentTo: { faceNormal: "+z" } })
  .expect({ count: 4 })                                  // a build error, not a surprise
  .fillet(0.8);

You never name an edge by index. You describe it — the circular rims that open onto the top face — and the description still works after you move a hole.

One graph, one exact kernel

Your script builds a small JSON graph, and an exact kernel reads it — OpenCASCADE: real faces and edges, STEP export, and a refusal rather than an approximation when it cannot do something faithfully. The graph is what keeps today's script working after tomorrow's kernel swap.

Parts get measured, not assumed — volume, wall thickness, whether two bores actually meet. An agent gets the same numbers over MCP, with no screen to look at.

And when it does look, it gets pictures made for reading: views that share one scale, sections through the inside, and every surface coloured by the tag that built it.

It is aimed at small mechanical parts — brackets, flanges, manifolds, heat sinks, the things you print or machine one of. examples/ is twenty-two.

Why not CadQuery, build123d or OpenSCAD

Those are good, and much older. Three differences. Selectors are descriptive and checked — .expect({ count: 4 }) turns a selector that silently started matching three edges into a build error, which is the failure mode that makes code-CAD fragile. What you see, what you measure and what you export come off one exact B-rep, so the preview and the STEP file are one model, not two projects. And measurement is an output rather than something you eyeball, which is what makes the MCP surface real instead of a wrapper.

Install it

With Homebrew, on macOS or Linux (Linux from 0.0.6), which installs the host and no window:

brew tap ierehon1905/parcad
brew trust ierehon1905/parcad
brew install parcad

Then add it to your agent, which starts parcad when it needs it. To use the app yourself, run parcad serve and open — the whole app is there, in your browser.

On Windows, from 0.0.6, the desktop app through winget:

winget install ParCAD.ParCAD

Nothing on the Homebrew path is quarantined. The same parcad is the headless driver (parcad part.js), and a client of the running host: parcad tools lists what an agent can call over MCP and parcad call calls it from a shell.

Or the desktop app, from Releases:

Shortened. The full README is on GitHub.