FreeCAD MCP

Two ways to use this

This project exposes FreeCAD as a server that two different kinds of users talk to:

	Who	How	What for
MCP server (port 10944)	AI agents — Claude Desktop, Cursor, Continue, any MCP client	Your AI assistant calls tools like `step_to_stl()` or `cfd_run_solver()` directly. You type "convert this STEP file and slice it for my printer" and it does the whole chain.	Hands-free CAD work. Agentic loops: the AI iterates design → simulate → analyze → refine without you touching the GUI.
Web dashboard (port 10945)	Humans — you, in a browser	A dark-mode React app with pages for converting files, browsing models, configuring fluid simulations, chatting with an LLM about CAD, viewing logs, and changing settings.	Point-and-click workflow. Upload a file, click Convert, download the STL. Fill in a form, click Run, see simulation results.

Both talk to the same FreeCAD backend. An AI agent can prepare a simulation case, and you can inspect the results in the web dashboard — or vice versa.

Scripting & automation

Everything the AI agents and the web dashboard call goes through the same REST API at http://localhost:10944/api/v1/. You can script against it directly from Python, curl, or any HTTP client:

import requests

# Ask the AI assistant to convert a file — same API the webapp uses
r = requests.post("http://localhost:10944/api/v1/control/tool", json={
    "tool": "step_to_stl",
    "arguments": {"file_name": "bracket.step", "output_name": "bracket.stl"}
})
print(r.json())  # {"success": True, "output": "bracket.stl", ...}

# Or run a parametric study across 10 different geometries
for length in [0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 4.0, 5.0, 7.5, 10.0]:
    requests.post("http://localhost:10944/api/v1/control/tool", json={
        "tool": "cfd_create_domain",
        "arguments": {"domain_type": "channel", "length_m": length, "case_name": f"study_{length}m"}
    })

This is how you build headless pipelines: generate 100 geometry variants in a loop, run simulations on each, collect results, train a surrogate model. No GUI needed — FreeCAD runs in the background as a geometry engine.

What can I do with this?

FreeCAD is a free professional CAD modeler (like SolidWorks but open-source). Through the API, you can:

Category	What you can do
Mechanical parts	Convert STEP files to 3D-printable STL meshes. Create boxes, cylinders, spheres, cones by describing them. Inspect part volumes and bounding boxes. All managed through the persistent CAD file depot with full CRUD.
Architecture & construction	Create walls, floor slabs, columns (rectangular or steel H-sections), windows, doors, and sloped roofs — all in millimeters. Export to `.ifc` files that architects and structural engineers can open. Import `.ifc` files from other design tools. This is called BIM (Building Information Modeling) — 3D building design where every element knows what it is (a wall, a door, a load-bearing column), not just a shape. FreeCAD's Arch workbench handles this. QCAD is a 2D drafting tool — different domain.
Fluid simulation	Two CFD solvers, each with a different job: OpenFOAM (CPU, Docker) — complex geometry, full turbulence models (kEpsilon, kOmegaSST, LES), multiphase, heat transfer, 30-year industry standard. Use when you need fidelity or have complex geometry. FluidX3D (GPU, OpenCL) — Cartesian grids only, fast iteration, millisecond-per-timestep throughput on any GPU (NVIDIA, AMD, Intel, Apple). Use when you have a GPU and the geometry is simple enough to fit a uniform grid. Also: describe a flow problem in plain English and let an LLM write the solver config.
Structural analysis (FEM)	Run finite element analysis with CalculiX: stress, strain, displacement, safety factor. 10 material presets (steel, aluminum, titanium, carbon fiber...), automatic meshing via Gmsh, and an end-to-end `run_fem_analysis` convenience tool. Same tool Formula 1 teams use for chassis and wing analysis.
3D printing	Slice STL files into G-code for your printer. Configurable profiles for different printers, filaments, and layer heights.
Model marketplace	Search Printables, Thingiverse, and GrabCAD for ready-made parts. Import them directly into your workspace.
Machine learning	Export point clouds from fluid domains to train neural networks that replace slow physics simulations. Generate datasets by running parameter sweeps automatically.


Example use cases	"What's inside this STEP file?" — "Convert this assembly to 3D-printable STL" — "Design a room with two windows and a door" — "Simulate water flow through this pipe at 2 m/s" — "Find me a gear on Printables and slice it for my MK4" — "Train a neural network to predict airflow instead of running the full simulation every time" — "Browse my persistent CAD depot, create a test bracket, rename and tag it"
What it connects to	FreeCAD 1.1.1+ (the CAD engine), OpenFOAM 10 via Docker (fluid simulation), FluidX3D via OpenCL (GPU-accelerated simulation), PrusaSlicer 2.8+ (3D printing), Ollama (local LLM), Printables / Thingiverse / GrabCAD (model search). ParaView (optional, for professional CFD viz): `winget install Kitware.ParaView`
Ports	10944 = MCP server (AI agents), 10945 = web dashboard (humans), 10946 = FreeCAD bridge (internal)
Start	`just bootstrap` then `start.ps1`

Documentation Index

Guide	Content
Installation	Prerequisites, FreeCAD setup, PrusaSlicer, `just bootstrap`
Architecture	TCP bridge, subprocess fallback, port layout, file pipeline
MCP Tools	All 46 tools with examples: geometry conversion, architecture (walls/floors/roofs/IFC), structural FEM (CalculiX stress/strain), fluid simulation (OpenFOAM + FluidX3D), 3D printing, marketplace search, LLM assistance, and CAD file depot with persistent storage + full CRUD
Fleet CFD pipeline	qcad DXF → freecad OpenFOAM/FluidX3D → resonite/godot/robotics handoff
Fluid simulation guide	Full walkthrough: creating domains, configuring physics, setting boundary conditions, running solvers, reading results, parametric sweeps, natural language config, and exporting data for neural network training
Flow visualization & render guide	What FluidX3D produces (VTK, OBJ, WebM, PNG, CSV). How to turn it into video, game engine assets, or VR visuals. Glossary of terms (VTK, OBJ, streamlines, LBM, MLUPS, heatmaps). Integration with Unity3D, Godot, Blender, Resonite.
OpenFOAM & GPU solvers	Solver reference, turbulence model guide, GPU acceleration options (FluidX3D on RTX 4090 and Apple Silicon), Mac vs PC comparison
AI Tooling	Ollama chat, agentic CAD reasoning, sampling workflows
About FreeCAD	History, community, Python scripting, 300+ workbenches, vs SolidWorks/Fusion
Marketplace	Searching and importing models from Printables, Thingiverse, GrabCAD
Webapp README	Dashboard frontend: pages, viewer, proxy, development

Quick Start

just bootstrap   # uv sync + npm install
start.ps1        # kills zombies, starts backend + frontend, opens browser

MCP Client Config

{
  "mcpServers": {
    "freecad": {
      "url": "http://localhost:10944/sse",
      "transport": "sse"
    }
  }
}

Quality

Python: Ruff linter + formatter (zero-config, sub-millisecond)
Frontend: Biome linter + formatter
Task runner: just — just lint, just fix, just dev, just tauri-build
Native desktop: Tauri 2.0 + PyInstaller sidecar — just tauri-build (ports 10944/10945)
AI protocol: FastMCP 3.2 with SSE transport

CFD fleet moat

Only fleet repo with OpenFOAM + GPU FluidX3D + NL2FOAM + PINN sampling + VTK/OBJ export for downstream Resonite/Godot/robotics handoff. Competitive analysis: mcp-central-docs/projects/CAD_FLEET_COMPETITIVE.md. Pipeline: docs/fleet-pipeline.md.

just fleet-e2e-offline   # CI-safe config chain smoke
just fleet-e2e-integration   # GPU FluidX3D setup->compile->run (local)
just fleet-e2e-chain     # qcad extrude -> freecad setup (HTTP)

License

MIT — see LICENSE.

Repo	What it does
QCad MCP	2D technical drafting — floor plans, schematics, blueprints
Blender MCP	3D modeling, animation, rendering, sculpting
Unity3D MCP	Game engine, real-time 3D, physics, VR/AR
Inkscape MCP	Vector graphics — logos, illustrations, SVGs
GIMP MCP	Raster image editing — photos, textures, pixel art

Freecad MCP

At a glance

README