Rupa3D MCP server

Rupa3D

3D tooling where every asset carries its own proof.

Rupa3D builds 3D — exact b-rep CAD (OpenCascade), meshes (headless Blender), and scenes with physics (three.js + Rapier) — and then measures what it built against closed-form references or independent measurements. The measurements travel with the asset: a certificate embedded in the GLB at asset.extras.rupa3d, readable with a few lines of code and no Rupa3D installed.

It is designed to be driven by AI agents over MCP, and by CI over a command-line gate.

Bahasa Indonesia

Status — 1.6.2 · 297 tests · MIT · on npm as

rupa3d · in the official MCP

Registry as io.github.tiranyx/rupa3d.

Developed and tested on Windows 11 with Node 22. Tool descriptions and

messages are still in Indonesian.

Use it as an MCP server

claude mcp add rupa3d -- npx -y rupa3d

Or in any MCP client configuration:

{ "mcpServers": { "rupa3d": { "command": "npx", "args": ["-y", "rupa3d"] } } }

Requires Node ≥ 22.14. Workspaces live in ~/.rupa3d; override with RUPA3D_RUANG.

Blender is optional. 23 of the 36 tools run without it:

The other 13 drive headless Blender — bpy modeling, import (GLB, OBJ, FBX, STL, PLY, SVG), renders, LOD, baking, collision proxies, source-mesh topology. Blender is found under C:\Program Files\Blender Foundation\, at RUPA3DBLENDER, or as blender on PATH.

What "proves" means

Every number below comes from the test suite or the documentation in this repository.

(area from the shoelace formula), revolution against Pappus's theorem. Recorded errors range from 0 % to 6.03e−14 %; the tests fail above 1e−10 % for extrusion and 1e−8 % for revolution. A result that misses its reference is never saved.

• Sketch → solid. Extrusion volume is checked against area × distance

analytic value; the best mesh tessellation tried (9,252 triangles) is off by −0.0016 %. CAD export reports which of the two you are holding (eksak) and by how much the approximation misses.

• Exact vs approximate. For a Ø140 mm flange the b-rep volume equals the

half the smallest edge came back ok with a broken solid whose volume grew by 1,971 mm³ on a 6,000 mm³ box. Independent checks reject it.

• The kernel's ok: true is not accepted as proof. A fillet larger than

30, 60 and 240 Hz, and the error shrinks as the step shrinks — the test guards the direction, so it cannot pass on a lucky timestep.

• Physics. Free-fall time matches h = ½gt² within one simulation step at

certificate as its uncompressed twin: 2,208 triangles, 0 failures, 6 warnings.

• Compressed files. A Draco-compressed sphere receives the same

accessor counts or oversized skins are rejected with a message naming the defect, instead of hanging or passing. Each case runs in a child process with a time limit, so a regression fails the suite instead of freezing it.

• Hostile files. GLBs with node cycles, out-of-range indices, forged

Command-line gate

npx -p rupa3d rupa periksa model.glb

Exit code 0 — passes every required rule. 1 — fails a required rule (fix the asset). 2 — could not be checked (fix the tool or the input). --json for machine output, --spek for another spec, --tempel to write a certified copy.

Shortened. The full README is on GitHub.