Skip to main content
Glama
graphgrow

root-ext-cad

by graphgrow

root-ext-cad

CAD-engineering MCP tool server — code-CAD builds (build123d/CadQuery), model inspection, meshing, DFM-lite validation, and mechanical calculators. Built as the mechanical-engineering pack for [ROOT Workspace], but it speaks plain MCP stdio and works with any MCP host.

The workspace client stays domain-free by doctrine; this repo is where the CAD knowledge lives. Tools reach both ROOT runtimes (the brain-governed deployment and the lite Tier-0 harness) through the client's mcp-bridge, and every call is governed — write-class tools gate like any write; network tools park for approval.

The charter (binding)

Source is truth; artifacts are derived. The pack never overwrites a user's file in place: geometry writes land in a caller-named output directory (default derived/ beside the source), and every write tool refuses an output path that collides with its input. tests/test_server.py::test_no_inplace_mutation_tools_exist and the collision-refusal tests enforce this in CI, not just in prose.

Parametric source (a build123d or CadQuery .py script) is the model; STEP/STL/glTF/3MF are exports. model_build executes scripts — it says so plainly, runs them in this pack's own venv as a resource-capped subprocess, and is classed write, never read.

Related MCP server: build123d-mcp

Tools

Tool

Class*

What it does

model_inspect

read

STEP/BREP/STL → bounding box, volume/area, mass (given density), topology counts, per-solid summary

model_validate

read

DFM-lite lint: open (non-watertight) shells, invalid B-rep, tiny edges/faces, pairwise interference & clearance — every finding carries its evidence and threshold

model_build

write

Run a build123d/CadQuery parametric script (sandboxed subprocess, CPU/memory/wall caps) → STEP/STL/glTF/3MF into the output dir + a build report

model_mesh

write

STEP/BREP → tessellated STL/glTF/3MF with linear + angular tolerance controls

calc_fits

read

ISO 286 limits & fits (hole-basis subset), clearance/interference verdicts

calc_fastener

read

Metric coarse threads: pitch, stress area, tap & clearance drills, preload/torque first-cut

calc_beam

read

Beam bending first-cut: cantilever / simply-supported, point / UDL, rect / round / tube sections

calc_mass_rollup

read

Assembly mass + 3D center of gravity from a component list, envelope check

model_diff

read

Geometric diff of two revisions: volume/area/bbox deltas, bodies unchanged/moved/added/removed by shape signature — geometry, not a feature tree, and the report says so

drawing_project

write

Named-view (front/top/…/iso) SVG or DXF projection: visible edges + dashed hidden layer — a silhouette for hand-off, not a dimensioned drawing

text_to_cad

external

Starter model from a text prompt via Zoo's ML-ephant API (+ its editable KCL source when returned); needs ZOO_API_KEY, refuses by name without it

* Classes derive fail-safe from annotations in the ROOT client: read tools declare readOnlyHint: true, openWorldHint: false; write tools declare readOnlyHint: false, openWorldHint: false (the closed-world write class); network tools (Phase CAD-D: text_to_cad) will declare openWorldHint: true and refuse by name without their key.

Connect it to ROOT Workspace

Settings → Connected tools → the connect form:

  • Name: cad-eng

  • Command: uv run --project /path/to/root-ext-cad root-ext-cad (a local checkout), or once published: uvx root-ext-cad

Run the row's Check (doctor) after connecting. Keep model files and build scripts inside granted folders so the agent can cross-read them with fs.*. Builds must finish inside the bridge's call ceiling (120 s; 90 s once promoted) — the default subprocess wall cap is 75 s.

Development

uv venv .venv
uv pip install --python .venv/bin/python -e ".[dev]"
.venv/bin/pytest

Layering rule: server.py is wiring only; every capability lives in a plain-Python module (inspection.py, meshing.py, validate.py, buildrun.py, calc.py, geometry.py) with logic separated from I/O so the rules stay unit-testable on synthetic shapes — no fixture model files required.

Roadmap

  • Later, each behind its own ADR: FreeCAD GUI-bridge companion, Onshape connector, KCL as a second source dialect, FEA hand-off recipes, PMI/GD&T presence reporting

Install Server
F
license - not found
A
quality
B
maintenance

Maintenance

Maintainers
Response time
Release cycle
Releases (12mo)
Commit activity

Resources

Unclaimed servers have limited discoverability.

Looking for Admin?

If you are the server author, to access and configure the admin panel.

Related MCP Servers

  • F
    license
    -
    quality
    D
    maintenance
    MCP server that exposes CAD geometry reasoning over STEP files to LLMs, allowing natural language queries about parts, assemblies, dimensions, holes, and mass properties.
  • A
    license
    -
    quality
    A
    maintenance
    MCP server for interacting with Creo Parametric CAD software. Enables tool calls like file_open, feature_set, and J-Link execution through a Python-based server and optional Java gateway.
    MIT
  • A
    license
    -
    quality
    C
    maintenance
    Enables AI agents to programmatically build, analyze, and export 3D CAD geometry using FreeCAD through REST or MCP tools.
    1
    MIT

View all related MCP servers

Related MCP Connectors

  • MCP server for generating rough-draft project plans from natural-language prompts.

  • MCP server exposing the Backtest360 engine API as tools for AI agents.

  • A comprehensive Model Context Protocol (MCP) server that enables AI assistants to interact with yo…

View all MCP Connectors

Latest Blog Posts

MCP directory API

We provide all the information about MCP servers via our MCP API.

curl -X GET 'https://glama.ai/api/mcp/v1/servers/graphgrow/root-ext-cad'

If you have feedback or need assistance with the MCP directory API, please join our Discord server