OpenSCAD MCP Server

A Model Context Protocol (MCP) server that lets AI assistants design 3D-printable parts and assemblies in OpenSCAD: render with a stated scale, measure exact geometry, check assemblies for interference and clearance, extract holes and features, judge printability, and export. Built with FastMCP for Python; OpenSCAD 2021.01 is the supported floor and dev snapshots are used when present.

Prerequisites

• OpenSCAD installed on your system

• uv (recommended) or Python 3.10+

Related MCP server: 3D MCP Server

Installation

The server is published on PyPI as openscad-mcp, so uv runs it with no clone and no virtualenv: uvx openscad-mcp. uv keeps a cached copy; uv tool upgrade openscad-mcp (or uvx openscad-mcp@latest) pulls a new release, and uvx openscad-mcp@0.6.1 pins one.

Claude Code

Add the server with a single command:

claude mcp add openscad --transport stdio -- uvx openscad-mcp

Or, if OpenSCAD is not on your PATH:

claude mcp add openscad --transport stdio \
  --env OPENSCAD_PATH=/path/to/openscad -- uvx openscad-mcp

Use the --scope flag to control where the configuration is saved:

The repository is also a Claude Code plugin (skill plus server): /plugin marketplace add robertcoop/openscad-mcp then /plugin install openscad-mcp@openscad-mcp.

Claude Desktop

Add to your configuration file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "openscad": {
      "command": "uvx",
      "args": ["openscad-mcp"],
      "env": {
        "OPENSCAD_PATH": "/usr/bin/openscad"
      }
    }
  }
}

Then restart Claude Desktop.

Cursor / Windsurf / VS Code

Add a .mcp.json file to your project root:

{
  "mcpServers": {
    "openscad": {
      "command": "uvx",
      "args": ["openscad-mcp"]
    }
  }
}

Manual / Standalone

# From PyPI (no install required)
uvx openscad-mcp

# The development version, straight from GitHub
uvx --from git+https://github.com/robertcoop/openscad-mcp.git openscad-mcp

# Or clone and run locally
git clone https://github.com/robertcoop/openscad-mcp.git
cd openscad-mcp
uv run openscad-mcp

# Run an assembly check file from a shell or a Makefile (exit code 0/1/2)
uvx openscad-mcp check checks.yaml --allow /path/to/project

Available Tools

Every tool response carries errors, warnings and hints parsed from OpenSCAD's output. Check them: OpenSCAD exits 0 on a failed assert() or an unknown module and draws a blank scene.

Rendering

Assemblies

Parts are given inline as parts=[{name, code, place?, frame?, ghost?, mass_g?, motion?}] or in a check file (frames, quality, parts, checks, model). Any number in a rule may be a SCAD expression string (point: "[BOLT_R, 0, BASE_H]") evaluated in the model's scope, so checks follow the parameters rather than a copy of them. openscad-mcp check <file.yaml> runs a check file from the shell with a meaningful exit code, so make check is one call.

Export & Model Management

Measurement & Validation

Project Support

Usage Examples

Once connected, ask your AI assistant:

• "Render a cube with rounded edges"

• "Show me the front and top of this model with a scale bar"

• "What is the volume and are there any cavities?"

• "Cut a section through the lid at z = 12 and tell me the wall thickness"

• "Colour the body and lid differently and ghost the body"

• "What clearance should I use for an M3 screw and a press-fit 608 bearing?"

• "Compare the model before and after changing the radius to 15"

• "Export my gear model to STL"

The server also publishes MCP resources (openscad://conventions, openscad://cheatsheet, openscad://reference/{topic}) and server instructions with the coordinate and assembly conventions it expects. A Claude Code skill lives in skills/openscad-design/SKILL.md and the repo can be installed as a Claude Code plugin (.claude-plugin/).

Tool Parameters

API.md documents every tool and every parameter. Four parameters account for most of the questions:

• parts — a list of {"name": ..., "code": ..., "place": ...} objects, accepted by render(mode="parts"), measure(mode="parts"), check and export_model. code is the statement that instantiates the part (lid();) and place is an optional OpenSCAD transform wrapped around it (translate([0,0,20])). Each part is exported on its own, so its identity survives; an assembly is never unioned. measure(mode="parts") measures each part in its placed position (the response says frame: "assembly"), the same grammar check, render and export_model use.

• quality — draft, normal, high, or an integer $fn. It sets $fn/$fa/$fs for the run. It is a correctness knob, not only a speed one: check reports a distance smaller than the tessellation error bound as UNRESOLVED rather than guessing, and the fix is a higher quality.

• variables — a dict injected as OpenSCAD variables. In the wrapped modes (section, parts, scad_eval) they are appended to the wrapped module body rather than passed with -D, and they are injected at file scope as well, so a constant derived inside an included file still sees them. This is how you set a $preview guard variable for an export.

• include_paths — extra directories added to OPENSCADPATH. When MCP_ALLOWED_PATHS is set, every entry is validated against it, as is every file OpenSCAD actually reads.

Exactly one of scad_content or scad_file is required by every tool that takes source. All parameter parsers accept multiple input formats (JSON strings, lists, dicts, CSV) for AI assistant compatibility.

Each image costs roughly 640 vision tokens at 800x600, and image sizes are clamped to 1568 px on the long edge, above which vision models downscale anyway. Ask for the views that answer a question rather than all of them. Auto-fit renders (grounded=false) have no recoverable absolute scale, which the digest states.

Configuration

Environment Variables

YAML Configuration

Create a config.yaml for advanced configuration:

server:
  name: "OpenSCAD MCP Server"
  version: "0.1.0"
  transport: stdio

rendering:
  max_concurrent: 5
  timeout_seconds: 300
  default_color_scheme: Cornfield

cache:
  enabled: true
  max_size_mb: 500
  ttl_hours: 24

security:
  rate_limit: 60
  max_file_size_mb: 10
  allowed_paths:          # null = no path validation at all (a warning is logged)
    - /home/me/projects/parts
  max_memory_mb: 4096

Security

Threat model

The server runs OpenSCAD on source it is handed. OpenSCAD can read any file the process can read, through include <>, use <>, import() and surface(), and can return what it read as echo output or as geometry. The guarantees below hold only when allowed_paths is configured. Out of the box it is unset, no path validation is performed, and the server logs a warning at startup saying so.

What is enforced:

• Path validation on arguments: scad_file, include_paths (in every tool) and export output_path must lie inside allowed_paths. Containment uses resolved paths, so symlinks and .. cannot escape.

• Path validation on the dependency closure: every file OpenSCAD actually read is recorded with -d and checked after the run. If any lies outside allowed_paths, the standard library directories, or the server temp dir, the output (image, mesh, echo text) is withheld and the call fails. This closes the include <...>-as-data and surface(file=...) channels.

• Memory ceiling: each OpenSCAD process runs under RLIMIT_AS (max_memory_mb, default 4 GB) on POSIX hosts. OpenSCAD has no ceiling of its own; a small minkowski() can otherwise consume all host memory.

• Timeout: timeout_seconds, default 300 s; partial stderr is kept.

• Echo channel bounds: echo_output is capped (200 lines, 2000 chars per line) and labelled as untrusted content from the rendered file.

• File size limits, variable name validation (^\$?[a-zA-Z_][a-zA-Z0-9_]*$) and model name validation (no path traversal) as before.

What is not enforced: no OS-level sandbox (no network isolation, no filesystem namespace). For untrusted input run the server inside a container or under Landlock/bubblewrap with only the project directory mounted.

Why --hardwarnings is off

--hardwarnings stops OpenSCAD at the first warning but still exits 0, so it produced blank renders and silently truncated echo_output with no indication. Warnings now reach the assistant through the structured warnings, errors and hints fields on every tool response instead. Set MCP_HARD_WARNINGS=true to restore the flag.

Development

# Clone the repo
git clone https://github.com/robertcoop/openscad-mcp.git
cd openscad-mcp

# Install dependencies (the dev extra; there is no dependency-groups table,
# so `uv sync --dev` would remove pytest, ruff, black and mypy)
uv sync --extra dev

# Run the server
uv run openscad-mcp

# Run the assembly checker on a check file
uv run openscad-mcp check examples/checks/turntable.yaml

# Run tests
uv run pytest

# Lint & format. The tree carries pre-existing findings, so expect noise;
# CI gates only on `ruff check --select F,E9,B src/openscad_mcp/`.
uv run ruff check src/ tests/
uv run black --check src/ tests/

# Type check (also not clean today)
uv run mypy src/

Project Structure

openscad-mcp/
├── src/openscad_mcp/
│   ├── server.py            # FastMCP server, the 12 tools, rendering, cache, CLI
│   ├── assembly.py          # Part/Frame/Assembly model and the check-file grammar
│   ├── checks.py            # RuleEngine: the rules a check file can ask for
│   ├── geom.py              # Mesh kernel: BVH, tri-tri distance, winding number, sweeps
│   ├── csgfeatures.py       # CSG-dump parser: holes, bosses, cross-part alignment
│   ├── massprops.py         # Mass, centre of mass, inertia by tetrahedra
│   ├── printability.py      # Overhangs, wall thickness, islands, orientation candidates
│   ├── analysis.py          # Static analysis: BOSL2 $var shadowing lint, constant tracing
│   ├── parts_catalog.py     # Purchased-parts catalog loader and self-check
│   ├── parts/*.scad         # One generated BOSL2 file per catalogued part
│   ├── threemf.py           # Multi-object 3MF writer
│   ├── wrappers.py          # Source-level wrappers: include hoisting, variable injection
│   ├── diagnostics.py       # stderr -> Diagnostics; -d deps parsing; repair hints
│   ├── camera.py            # Orthographic camera model, fit, annotation, spatial digest
│   ├── mesh.py              # Stdlib STL/SVG analysis: welding, components, volumes
│   ├── reference.py         # Sourced engineering data: fits, fasteners, inserts, DFM
│   ├── types.py             # Pydantic models and enums
│   └── utils/config.py      # Configuration with env/YAML/dotenv support
├── tests/                   # ~1,500 tests; 80% coverage floor
├── evals/                   # Deterministic geometry eval harness, 15 tasks
├── examples/checks/         # A worked check file and its model
├── skills/openscad-design/  # Claude Code skill: the design loop
└── .claude-plugin/          # Claude Code plugin manifest

Testing

# Run all tests with coverage (about 1,500 tests, roughly a minute)
uv run pytest

# Run specific markers: unit, config, integration, slow, performance, edge, render
uv run pytest -m unit
uv run pytest -m performance
uv run pytest -m "not slow"

# Run a single file, without the coverage gate
uv run pytest tests/test_helpers.py -v --no-cov

Most tests mock the OpenSCAD subprocess, so no OpenSCAD installation is needed to run the suite. A mock that stands in for a render has to write both output files OpenSCAD would have written: the -o target and the -d dependency file that the cache manifest is built from. Tests that do need the real binary skip themselves when it is absent.

CI runs the suite on Python 3.10 and 3.12 with OpenSCAD 2021.01 and BOSL2 installed, under xvfb-run because PNG export on 2021.01 needs a display. It also builds the wheel, installs it in a clean environment, and asserts that a client sees exactly 12 tools. Coverage floor: 80%.

Troubleshooting

OpenSCAD Not Found

# Check if OpenSCAD is installed
which openscad        # Linux/macOS
where openscad.exe    # Windows

# Set the path explicitly
export OPENSCAD_PATH=/path/to/openscad

Server Not Connecting

# Verify the server starts correctly
uvx openscad-mcp

# In Claude Code, check MCP status
/mcp

Render Timeout

Increase the timeout:

export MCP_RENDER_TIMEOUT=600

"Path not allowed" / "outside the allowed paths"

MCP_ALLOWED_PATHS is set and the file is not under one of its roots. The check covers arguments and, separately, every file OpenSCAD actually opened, so a model that lives inside an allowed root but does include <../shared/lib.scad> outside it is refused and the output is withheld. Add both roots:

export MCP_ALLOWED_PATHS="$HOME/cad:$HOME/cad-shared"

Leaving the variable unset disables path validation entirely, which the server logs a warning about at startup. See Threat model.

An export or a measurement comes back empty

The usual cause is a $preview guard: a file that instantiates its geometry only inside if ($preview) renders in the GUI and exports nothing, because $preview is false for a render to a file. Pass the guard variable explicitly:

measure(scad_file="part.scad", variables={"$preview": true})

A difference() whose first child is smaller than what follows also yields nothing. validate(mode="syntax") will not catch either one; the exit code is 0 in both cases.

check rows say UNRESOLVED

The distance in question is smaller than the tessellation error of the mesh, so the answer would be an artefact of $fn rather than of the design. Re-run with quality="high" or an explicit integer $fn. Every row reports the quality.fn it was computed at.

A result looks stale

Renders are cached under ~/.cache/openscad-mcp/, keyed on every render parameter plus a manifest of every file OpenSCAD read. Editing an included file normally invalidates the entry, but if a response says cached: true and the number disagrees with the source, clear it:

clear_cache()

Renders fail on a headless machine

OpenSCAD 2021.01 needs a display to export PNG even in headless mode. Run the server under a virtual framebuffer:

xvfb-run -a uv run openscad-mcp

Exports, measurements and checks produce meshes rather than images and do not need a display.

Contributing

1. Fork the repository

2. Create a feature branch (git checkout -b feature/my-feature)

3. Make your changes with tests

4. Ensure tests pass (uv run pytest)

5. Open a Pull Request

Commit style: feat:, fix:, docs:, refactor:, chore:

CONTRIBUTING.md covers the module map, how the tests mock OpenSCAD, the design rules that should not be undone, and how to add a parts-catalog entry or a check rule.

License

MIT — see LICENSE

Acknowledgments

• OpenSCAD — Programmable CAD software. Everything here is a wrapper around its CLI.

• FastMCP — Python MCP framework

• Model Context Protocol — The MCP specification

• BOSL2 — Used, not vendored, by the purchased-parts catalog and by the BOSL2 anchor probe. BSD-2-Clause; install it separately.

• Dimensional data in reference and parts is cited entry by entry, with a confidence label on every number. See src/openscad_mcp/parts/README.md.