3D CAD BasePlate Generator MCP.

A NitroStack MCP server for the bottle-cap-cad workflow: turn a photo of a broken bottle cap, worn plug, or cavity into a printable replacement.

Vision/dimension extraction from the source photo happens client-side (in the MCP client, e.g. Claude) — none of these tools re-derive dimensions from pixels. generate_visual_mesh is the one exception, since Meshy infers geometry directly from the image itself.

Tools

A fill-geometry-guide MCP prompt is also registered, walking a caller through the fields fill expects.

Suggested pipeline: generate_visual_mesh (rough visual reference from a photo) → repair_mesh (make it manifold/watertight) → generate_precise_cap or fill for the actual dimensionally-accurate part → repair_mesh again if the CAD output needs cleanup after boolean operations.

Related MCP server: bottle-cap-mcp

Setup

npm install
pip install -r requirements.txt --break-system-packages   # or your own venv
cp .env.example .env   # set MESHY_API_KEY
npm run build
npm start

For local iteration without a build step: npm run dev.

generate_precise_cap, fill, and repair_mesh shell out to a local Python interpreter (python-runtime.ts auto-detects python3/python/py, or respects a PYTHON_EXECUTABLE override in .env) — Python 3 with the packages in requirements.txt needs to be installed wherever this server runs. generate_visual_mesh only needs MESHY_API_KEY; it makes no Python calls.

Structure

3D-CAD-Modeling-using-MCP/
├── src/
│   ├── app.module.ts          # registers the four tool providers + prompt
│   ├── main.ts                 # bootstrap / entrypoint
│   ├── index.ts                # re-exports main.ts
│   ├── modules/bottle-cap-cad/
│   │   └── bottle-cap-cad.prompts.ts   # fill-geometry-guide prompt
│   └── tools/
│       ├── shapes.ts            # shared circle/rectangle/stadium/polygon/freeform zod schema
│       ├── cap.tools.ts         # generate_precise_cap
│       ├── fill.tools.ts        # fill
│       ├── repair.tools.ts      # repair_mesh
│       ├── mesh.tools.ts        # generate_visual_mesh (pure TS, calls Meshy)
│       ├── python-runtime.ts    # local python3/python/py auto-detection
│       └── cad-api-client.ts    # optional HTTP client for a remote CAD API
├── cad/
│   ├── precise_cap.py           # CadQuery threaded-cap builder
│   ├── fill.py                  # CadQuery plug builder
│   └── repair_mesh.py           # PyMeshLab repair pipeline
├── package.json
├── tsconfig.json
└── requirements.txt

Notes on the current code

• Thread generation: precise_cap.py builds a solid, closed-top cap with a blind bore (not a through-hole) and sweeps one or more real helical thread ridges onto the skirt's inner wall along a cq.Wire.makeHelix() path. threadStarts (1–4), threadDepth, and topThickness are optional fields on generate_precise_cap; threadPitch falls back to an approximate PCO 1810-style pitch (3.18mm, single start) when omitted. This is a parametric approximation tuned for FDM printing, not a certified thread-spec generator.

• fill shapes: shapes.ts supports five top/bottom shape types — circle, rectangle (with optional corner radius), stadium (true rounded slot with semicircular ends), polygon (straight-edged, arbitrary point count), and freeform (a smooth closed spline through 6–300 traced boundary points, for organic/irregular openings). capStyle: "flush" is only implemented for circle/rectangle/stadium/freeform tops — a polygon top with flush is downgraded to none by fill.tools.ts, with a warning returned to the caller rather than failing silently.

• Low-confidence inputs: both generate_precise_cap and fill accept a confidence field (high/medium/low, from the caller's own vision-based dimension extraction). A low value still generates output, but the response includes a warning recommending the dimensions be confirmed before printing.

• Remote CAD API: cad-api-client.ts supports routing generate_precise_cap / fill / repair_mesh over HTTP to a CAD_API_URL instead of spawning Python locally, for hosting environments without a Python runtime. This repo does not currently include the server-side counterpart (no api/ directory) — CAD_API_URL / CAD_API_KEY are recognized but there's nothing to point them at yet. Leave both unset for local development; the tools fall back to the local-subprocess path automatically.

• Meshy polling: mesh.tools.ts polls Meshy's image-to-3d endpoint every 3s for up to 60 attempts rather than using a webhook callback. Meshy's free tier output is CC BY 4.0 licensed — check meshy.ai/pricing for current terms before relying on it beyond a demo.

• app.module.ts registers a harmless placeholder for NitroStack's OAUTH_CONFIG DI token purely to silence a noisy (but non-fatal) startup log from OAuthModule — this server has no HTTP/OAuth surface and the placeholder has no effect on transport selection.

Team

• Reshvanth Yeddla

• Nirlep Boddapally

• Tarun

• Vijay Reddy