fusionwright MCP server
Commit Oct 27, 2026Connects an AI assistant to Autodesk Fusion to automate 3D modeling, verification, rendering, and file exportation through a validation loop.
Also known as Cocoloco20/fusionwright
- Stars
- 2
- Tools
- 10
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Oct 27, 2026
- License
- MIT
What is fusionwright MCP?
fusionwright connects an AI assistant to Autodesk Fusion to automate 3D modeling, verification, rendering, and file exportation through a validation loop. It exposes 10 tools, such as fusion_status, fusion_api, fusion_run.
Summary generated from the project's README
About fusionwright MCP server
Automating 3D modeling and mechanical verification within Autodesk Fusion is what fusionwright enables for engineers and designers. By connecting an AI assistant like Claude or Cursor to a live Fusion session, users can describe complex parts and have the assistant write, execute, and validate the necessary Python code. This setup works on macOS or Windows using a local bridge add-in and requires Python 3.10 or higher.
Unlike simple scripting bridges, this server implements a validation loop to ensure accuracy. It uses tools like fusion_run to execute code, fusion_check to verify volumes and interferences against the Fusion kernel, and fusion_look to capture multi-angle screenshots. The system can also perform photoreal renders via fusion_render and export files in formats such as STEP, STL, or OBJ. This approach allows the AI to identify and fix mistakes in geometry before the final design is completed.
What you can do with it
- 1Generate and export enclosuresAsk the AI to create a specific PCB enclosure with vent slots and export it as a STEP file.
- 2Verify mechanical accuracyHave the assistant check a part for invalid bodies or interferences and repair them automatically.
- 3Automate aesthetic finishingRequest the addition of fillets to all outer edges and a specific material render.
- 4Validate geometry with arithmeticEnsure the generated 3D volume matches a specific calculated value using the Fusion kernel.
Written from the project's README. Every use case is checked against a quote from it.
fusionwright MCP tools 10
Detected from the READMEfusion_statusfusion_apifusion_runfusion_rollbackfusion_inspectfusion_checkfusion_lookfusion_renderfusion_exportfusion_ui_command
| fusion_status | — | |
| fusion_api | — | |
| fusion_run | — | |
| fusion_rollback | — | |
| fusion_inspect | — | |
| fusion_check | — | |
| fusion_look | — | |
| fusion_render | — | |
| fusion_export | — | |
| fusion_ui_command | — |
Automated checks
Ran Nov 14, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablecocoloco20/fusionwright is public and reachablegithub.com/cocoloco20/fusionwright ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-10-27github.com/cocoloco20/fusionwright ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model10 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNo tool named like delete, drop or execPass
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubfusionwright
Let an AI build in Autodesk Fusion for you.
An open-source MCP server that connects an AI assistant (Claude, ChatGPT, Cursor, anything that speaks MCP) to your live Fusion session. You describe what you want; the AI writes the Fusion code, looks at the result from several angles, measures and verifies it with Fusion's own kernel, fixes its mistakes, then renders it and exports files you can print, machine or send on. It is not tied to any kind of object: brackets, enclosures, gears, jigs, furniture, product mock-ups, anything solid.

examples/flange_fusion_run.py, run through fusion_run in a live Fusion document: sketch, extrude, join, cut. The script checks Fusion's volume (37.542 cm3) against the hand-computed value (37.542 cm3), then fusion_look took this picture.
What you can ask for
- "Make an 80 x 50 x 25 mm enclosure for a PCB with four screw posts and vent slots, then export a STEP and an STL."
- "Model an M6 flange bracket with a 4-hole bolt circle, check it, and show me from the front."
- "Take the part open in Fusion, find any invalid bodies or interferences and fix them."
- "Add a 2 mm fillet to every outside edge and render it in brushed aluminium."
- "Build a 24-inch bookcase from 3/4 inch plywood and give me every part as its own component."
The loop that makes it dependable: write code, look at it, measure it against the arithmetic, repair, repeat. Executable CAD is often not correct CAD (research), and a validation loop sharply improves results, so the tools are built around checking the work.
Requirements
- Autodesk Fusion on macOS (tested) or Windows (untested). The free Personal-use tier works.
- Python 3.10+ and an MCP client.
It only uses Autodesk's documented public API and a legitimately licensed Fusion; it does not touch licensing.
Quick start
-
Install the bridge add-in (once): run
./scripts/install_addin.sh, then in Fusion pressShift+S, pick FusionBridge, click Run (tick Run on Startup to make it automatic). -
Register the server with your client.
Claude Desktop (
claude_desktop_config.json):{ "mcpServers": { "fusionwright": { "command": "uvx", "args": ["--from", "git+https://github.com/Cocoloco20/fusionwright", "fusionwright"] } } }Claude Code:
claude mcp add fusionwright -- uvx --from git+https://github.com/Cocoloco20/fusionwright fusionwright -
Ask. Exports land in
~/fusionwright-projects/(override withFUSIONWRIGHT_HOME).
Tools (10)
A small surface on purpose: fewer, higher-level tools beat dozens of thin wrappers.
| Tool | Purpose |
|---|---|
fusion_status |
Is Fusion reachable? Which document is open? |
fusion_api |
Look up the real Fusion API (class members with docs) before writing code |
fusion_run |
Run Python inside Fusion (helpers predefined); returns a rollback checkpoint |
fusion_rollback |
Undo everything after a checkpoint |
fusion_inspect |
Compact design state: sizes, volume, invalid bodies, timeline problems |
fusion_check |
Kernel validation: body validity, timeline health, interference, expected size |
fusion_look |
Screenshots from iso / front / back / left / right / top / bottom |
fusion_render |
Photoreal render with Fusion's local renderer |
fusion_export |
STEP, STL, F3D, OBJ, 3MF, IGES, SAT (local files only) |
fusion_ui_command |
Run any of Fusion's ~3,000 UI commands by id, or search them |
Inside fusion_run the names app, ui, design, root, adsk, P3, VI, FO, IN, coll() and board() are predefined; assign your answer to result. Fusion works in centimetres internally.
Examples
| File | What it shows |
|---|---|
examples/flange_fusion_run.py |
sketches, extrude / join / cut, picking profiles by area, checking against hand arithmetic |
Contributions of more examples (enclosures, gears, sheet metal, assemblies, anything) are welcome.
How it compares
- Autodesk's official Fusion MCP (announced May 2026) is a thin scripting bridge: read / screenshot, run a script, undo / redo. fusionwright adds a verification layer (kernel checks, multi-angle images, rollback checkpoints, an API lookup, UI-command access).
- Community Fusion MCPs expose dozens of low-level CAD operations (the biggest lists 89 tools) and mostly stop there. fusionwright is deliberately small and runs its bridge on an authenticated loopback port.
Safety
See SECURITY.md. In short: loopback only, bearer token (~/.fusion-bridge-token, mode 0600), Host-header guard, request-size cap; fusion_run blocks file deletion, network access and cloud saves by default and returns a rollback checkpoint. It is a safety net, not a sandbox: code you let an AI run inside Fusion is powerful. Save your work first.
What's tested
- 9 unit tests: the add-in's HTTP layer with a stubbed Fusion (token, host guard, size cap, error tracebacks) and a real stdio MCP round trip (tool list, safety guard, error messages).
- Live tests against a real Fusion (
pytest tests/integration): build a part and verify sizes, volume and interference with the kernel, views that must differ from each other, 5 export formats, a real render, rollback, API lookup, UI-command search, and a full MCP client round trip that returns images. - A 30-case capability suite (
tests/integration/capability_suite.py) covering sketches, constraints, extrude / cut / revolve / sweep / loft, fillet / shell / chamfer / hole, patterns, booleans, components and joints, parameters, timeline edits, materials, measurement, 8 export formats, STEP re-import, CAM setup creation, configurations, rendering and more.
See docs/COVERAGE.md for what the Fusion API can and cannot do and docs/GOTCHAS.md for API traps found along the way.
Known limitations
- Fusion-native drawings (
adsk.drawing) can only be created from a design saved to your Autodesk cloud, so they are not automated here. Verify every dimension before cutting, printing or machining. - "Everything a human can do in Fusion" is not literally achievable through its API.
fusion_ui_commandandfusion_runcover a great deal, but some interactions (some drag gestures, cloud workflows) remain manual.
Roadmap
Adapter for Autodesk's official Fusion MCP as an alternative transport · more verified examples · driving models from Fusion user parameters · MCP Apps live 3D preview · MCP Tasks for long renders · a cross-model evaluation suite.
Not affiliated
fusionwright is an independent project, not affiliated with or endorsed by Autodesk, Anthropic or OpenAI. "Autodesk" and "Fusion" are trademarks of Autodesk, Inc.
MIT licensed. Contributions welcome.
fusionwright MCP questions
What tools does fusionwright MCP provide?
It exposes 10 tools, including fusion_status, fusion_api, fusion_run, fusion_rollback, fusion_inspect.
What operating systems are supported?
It is tested on macOS and is currently untested on Windows.
Does it require a paid Autodesk subscription?
No, the free Personal-use tier of Autodesk Fusion works with this server.
How does it ensure the AI's designs are accurate?
It uses a validation loop that writes code, inspects the design with Fusion's kernel, measures it against arithmetic, and repairs mistakes.
What file formats can be exported?
You can export files in STEP, STL, F3D, OBJ, 3MF, IGES, and SAT formats.