The three jobs Claude has to do
Abidi splits the pipeline into three stages. A mistake in any stage shows up as wasted filament.
- Modeling. An MCP server lets Claude drive your CAD tool. Abidi says there are dozens for FreeCAD, Blender, Fusion 360 and others.
- Slicing. The model has to become printer instructions (G-code), using a tested printer and filament profile.
- Handoff and monitoring. The sliced file has to reach the printer. Claude then watches the print, in Abidi's case through the printer's built-in camera.
His main caution is that connecting Claude to CAD and Home Assistant does not complete this pipeline. Slicing and file transfer each need their own piece.
Modeling: what the FreeCAD MCP setup involves
The public FreeCAD MCP project on GitHub shows how much is behind "install an MCP server". It is described as a way to control FreeCAD from Claude Desktop and other MCP clients, creating and editing models, running Python scripts, inspecting documents, and running FEM analyses. That matches what Abidi describes.
The project's quick start has two parts. An add-on runs inside FreeCAD, and an MCP server is launched by your client. To use it:
- Install the add-on and restart FreeCAD.
- Start the RPC server from the MCP Addon workbench.
- Add a server entry to Claude Desktop's configuration file.
- Keep FreeCAD open with the RPC server running.
Connections use localhost by default. If the add-on is not running, Claude has nothing to talk to.
Prompts need measurements
Abidi's example is a wall mount for a temperature sensor. "Make a sensor mount" leaves too many decisions open. A better prompt supplies:
- the sensor's measured dimensions
- the screw spacing
- the mounting orientation
- the cable clearance
- which parts of the case must stay accessible
He says Claude can get you most of the way and can nearly finish simple parts. Complex parts may still need manual tweaks in the CAD tool.
The Home Assistant link, and why the token matters
Abidi's printers are connected to Home Assistant through HACS plugins. He gave Claude a long-lived access token and it could talk to the printers. He adds that what you can do depends on the integration, so connecting a printer does not expose every operation.
Many people now wire Claude to Home Assistant this way. Several community guides describe the same pattern.
- One Home Assistant MCP project's Windows guide says to create the token under the profile's Security tab and copy it immediately, because it is shown only once.
- The same project's guide advises making a token just for this server, so it can be revoked separately without affecting anything else.
- Another guide suggests using a dedicated Home Assistant user, ideally non-admin, for the token. That would limit what a leaked or misused token can do.
Home Assistant's developer documentation says long-lived tokens are valid for 10 years. It also says the token string is not saved in Home Assistant, so you must record it somewhere secure. Treat it as a credential. It is not a harmless connection setting.
Abidi does not describe how he stores or scopes his token. Dedicated accounts, narrow permissions and revoking unused tokens are reader precautions from these guides, not safeguards he says he has in place.
Check your Bambu firmware first
This is the biggest compatibility catch for Bambu owners. Bambu Lab's blog post on its authorization-control firmware says that after the update, full control through Home Assistant is no longer possible. Home Assistant can still read some printer information, but certain functions are limited. The post says the new controls apply to X1 Series firmware 01.08.03.00 and above. It adds that all future Bambu Lab printer models will include authorization control as standard.
For anyone copying this workflow with a Bambu printer, this decides whether Claude can start jobs or only watch them. Abidi does not give his printer model or firmware version, so his experience can't be assumed to carry over to yours. Check what your printer's integration actually exposes before you plan around it.
Slicing: use known-good profiles
PrusaSlicer's wiki says the command-line interface loads profiles from AMF or 3MF files, and parameters given on the command line override the imported settings. It also notes that Windows users must use prusa-slicer-console.exe. That means a known slicing workflow can be wrapped in a script that Claude calls.
Abidi recommends a tested printer-and-filament combination. Don't ask the model to invent temperatures and machine settings for each job. He says supports and print orientation are where AI most often goes wrong, and that multi-color printing takes a lot of trial and error. For simple brackets and replacement parts, he found AI-assisted slicing surprisingly capable.
Getting the sliced file to the printer needs its own implementation. A narrowly scoped tool or Home Assistant script can do it, but not every printer integration provides this. Abidi gives no commands, profile names or end-to-end configuration, so this part is for you to build.
Security: Claude can act, not just advise
An AI assistant that can edit CAD files, run Python and control a heated machine has real capability. Anthropic's help documentation on custom connectors makes the same point.
- Custom connectors can reach services Anthropic has not verified, and once connected Claude can access those services and take action in them.
- The guidance says to connect only to trusted servers, review requested permissions, and watch for prompt injection and unexpected changes in tool behavior.
- It also says to review tool approval requests carefully and only click "Allow always" for a server and tool you trust to run unsupervised.
- You can revoke access by disconnecting the connector or changing permissions at the third-party service.
Local MCP servers configured in Claude Desktop's configuration file are a separate mechanism from remote connectors. Anthropic's documentation says they use your local network, which is how the localhost FreeCAD setup works.
In practice, keep the first runs supervised. Don't approve "always allow" for anything that can start a heater or move an axis. Disable tools you don't need for the current job.
What Claude can't do
Abidi is blunt about the physical side:
- Claude can't tell that your build plate is in upside down.
- It can't clear off something already sitting on the bed.
- Wall thickness, overhangs and print orientation still decide whether a part holds together.
- Claude needs real measurements and a description of the intended load, not permission to guess.
His advice is to treat your first versions as prototypes. Check the fit, report what needs changing, and revise both the design and the workflow. Nothing in his article suggests AI-made brackets are tested or safe for critical loads, and you shouldn't assume they are.
Practical takeaways
- Plan for three integrations, not one. CAD, slicing and printer handoff each need separate setup.
- Check printer compatibility first. Bambu's firmware change shows that a Home Assistant connection can be limited to monitoring.
- Lock down the token. Use a dedicated user and token, store it securely, and revoke it when you stop using it.
- Keep a human in the loop. Review tool calls and confirm before any physical job starts.
- Supply measurements and load requirements. Give Claude real dimensions, and treat early prints as prototypes.
The honest summary is that this is a hobbyist integration project, not a one-prompt feature. If you're comfortable with config files, tokens and a few failed prints, it can cut the gap between idea and test print. If you want a button that makes parts, it isn't there yet.