SolidWorks CAD Assistant & MCP Server

Languages: English | Español

Python MCP server for SolidWorks automation with 154 tools, plus an optional agent/prompt-testing layer for AI-assisted workflows.

See it work

Make a 100 x 50 x 10 mm AISI 1020 mounting plate with four M5 clearance holes, 10 mm in from each edge, and draw it fully dimensioned. Stack two of them 10 mm apart, check the result, and draw the assembly with a BOM.

The 58 tool calls an agent made from that prompt, run against SolidWorks 2026 through the same MCP tools a client calls, and recorded (sped up five times). Each result is checked against a hand calculation: the plate's volume and mass, the 10 mm gap and the BOM quantity. This is a replay of the calls, not a recorded model session.

Note: the demo is not perfect, and is not the most efficient way to get this result. The run pauses after each call so it can be read, and the calls are chosen to show what the tools can do: every sketch dimension is added one at a time, and the mate distance is stepped down only so the plates are seen coming together.

The walkthrough in the docs has the agent's plan and every call with its arguments. Run it yourself with the README demo; the calls are also in plate_stack_calls.json.

Overview

Tried it? Tell us what worked and what didn't in a feedback issue. Reports from real setups decide what gets fixed next.

This project focuses on practical SolidWorks automation with an AI-friendly loop:

1. describe intent
2. generate a plan
3. execute MCP tools
4. inspect results
5. iterate

It includes:

• core MCP runtime for SolidWorks tool execution
• COM/VBA routing and adapter safety wrappers
• tool coverage across modeling, sketching, drawing, analysis, export, automation, templates, and macros
• optional agent orchestration/testing utilities under src/solidworks_mcp/agents/

Demos

Each GIF is a live test driving SolidWorks 2026 through the MCP tools, with the call and its result captioned beneath. Click one to see what it checks, a prompt that asks for it, and its tool calls.

verify_model passes a block and names what is wrong	An assembly: parts, mates and interference	A cut reports the volume it removed
A revolve reports the volume it added	A round part's size is exact	Reading the model keeps the selection
A full drawing package from one call	Setting the sheet scale	A BOM on an assembly drawing
Each model dimension shown once	Clearing every overlap	Material and mass
Listing dimensions and measuring a face	Flat pattern of a sheet-metal part	Custom properties

See every demo in the docs: what each one shows, its prompt, every tool call with its arguments (all 58 for the demo at the top), and how to run them yourself.

Supported Today

• Windows + SolidWorks COM automation for the main CAD lifecycle.
• Tested on SolidWorks 2026. Earlier versions may work but are untested; if you try one, please open an issue with what worked.
• Modeling, sketching, drawing, analysis, export, automation, templates, and macro tools.

Not Yet / Simulated

• Mock adapter output is simulated and should not be treated as engineering truth.
• Live 3D viewport streaming in a UI.
• Checkpoint-level interference validation.
• Simple simulation and fluid workflows
• Simple Topology Optimization via SimulationXpress etc.

Install

You need Windows 10 or 11 and SolidWorks (opened at least once). In PowerShell, install uv, then open a new PowerShell window and run setup:

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
uvx solidworks-mcp-python setup

uv fetches Python 3.13 and the server from PyPI by itself; nothing to clone. The guided setup then:

1. finds your SolidWorks version and the MCP clients on your PC: Claude Desktop, Claude Code, VS Code, Cursor and LM Studio;
2. lets you pick which clients to connect;
3. starts the server once to check that it lists its tools (in mock mode, so SolidWorks isn't touched);
4. adds a solidworks entry to each client you picked, then tells you what to restart.

Nothing needs a login or an API key, and the server runs only on your PC. Setup changes only the solidworks entry in each client's config. It backs the file up first and never shows or changes your other servers. Other entries that start this server the old way (run-mcp.ps1 and the like) are offered for removal, so tools don't load twice.

To check it worked: open SolidWorks, restart your client, and ask "What model is open in SolidWorks?"

The package is solidworks-mcp-python. uvx solidworks-mcp installs a different project. To work on the server itself, install from a clone. More fixes are in Troubleshooting.

Connect your client

Run setup again whenever you want to add a client:

uvx solidworks-mcp-python setup

To set up a client by hand instead, each one runs uvx solidworks-mcp-python. If a client can't find uvx, use its full path, from (Get-Command uvx).Source.

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

claude mcp add --scope user solidworks -- uvx solidworks-mcp-python

{
  "servers": {
    "solidworks": {
      "type": "stdio",
      "command": "uvx",
      "args": ["solidworks-mcp-python"]
    }
  }
}

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

For mock mode, which needs no SolidWorks, use "args": ["solidworks-mcp-python", "--mock"]. Linux and WSL work for the docs, the tests and mock mode, but not for driving SolidWorks.

$env:SOLIDWORKS_MCP_SOC_LOGGING_ENABLED = "true"
$env:SOLIDWORKS_MCP_SOC_SESSION_ID = "my-bracket"
.\.venv\Scripts\solidworks-mcp.exe
# ...work in your MCP client...
.\.venv\Scripts\python.exe -m solidworks_mcp.agents.soc_exporter my-bracket my_bracket.py

Contributing

install.ps1 gives you an editable install, so code changes take effect the next time a client starts the server. For the development tools, install the extras:

.\.venv\Scripts\python.exe -m pip install -e ".[dev,test,docs]"

.\dev-commands.ps1 lists the common workflows:

• dev-install: install or update the local dev environment
• dev-test: run the standard test suite (the CI-safe subset)
• dev-test-full: run the full suite, including smoke and integration paths
• dev-lint and dev-format: lint checks and formatting
• dev-docs-build, dev-docs-strict and dev-docs-audit: build the docs, build them failing on warnings, or write an audit report to .generated/docs

To mirror GitHub Actions CI locally (Ubuntu, the conda env from solidworks_mcp.yml, and make test), run .\scripts\run-ci-local.ps1. The first run builds the image; add -NoBuild to re-run tests without rebuilding.

Docs

• Main docs site: https://andrewbartels1.github.io/SolidworksMCP-python/
• Home/overview: docs/index.md

Key docs sections:

• Getting Started: docs/getting-started
• MCP Server Guide: docs/user-guide
• Tool Catalog: docs/user-guide/tool-catalog
• Agents and Skills: docs/agents
• Planning/Roadmap: docs/planning

Direct links:

• Installation
• Quick Start
• Tutorial: U-Joint Assembly Build
• Tutorial Tracks
• VS Code MCP Setup
• Architecture
• Agents and Prompt Testing
• PydanticAI and Schemas

License

MIT License. See LICENSE.