Skip to content

FAQ

Short answers, each with a link to the page that has the details.

Getting started

Do I need to know CAD or Python? No. You describe the part and give a few measurements; the assistant writes the part script (a Python file in parts/<name>/), builds it, looks at it and checks it. Knowing a little helps you change a number yourself, but set_params and b123d-print set do that too. See How a session goes.

Which assistant should I use? Does it cost anything? b123d-print is free and open source (MIT). Any MCP client works; the best results come from a model that can see images, because the workflow has it look at its own renders and sections. The clients, what each supports, and how to connect it: Connect your MCP client. What the assistant costs is up to its provider.

The plugin, the MCP server or Docker: which one? They all run the same MCP server.

  • Claude Code plugin: one command in Claude Code, the server plus a skill, the project folder as workspace (Claude Code).
  • MCP server through uvx: for every other client (Claude Desktop, Cursor, Gemini CLI, Antigravity, ...): a few lines in the client's config (Quick start).
  • Docker: the strongest isolation (the way to get a real sandbox on Windows), or the shared HTTP server for several machines (Docker, Which mode?).

The install failed, or the first start times out. The first start downloads build123d, OpenCascade and the rest, which takes a while. Run uvx --from b123d-print b123d-print-mcp --help once in a terminal, then restart the client. If the client says uvx isn't found, use its full path. More in Troubleshooting.

Can I use it without an AI? Yes: the b123d-print CLI builds, checks, exports and makes templates from part scripts you write yourself (CLI and Python).

How do I update? It depends on how you installed it; for uvx, uvx --refresh ... once, then restart the client (Updating).

Parts and printing

What units? Millimeters, everywhere: the scripts, the checks and the files (3MF stores the unit; STL has none, and slicers read it as millimeters). Give inches if you like: the assistant converts them and confirms each value in mm (Measuring for the AI).

Why did it ask me ten questions before modeling? On purpose. A part is only as good as its measurements: the assistant lists every dimension it read, asks for the ones that decide the fit, and waits for you to confirm, rather than guess (Measuring for the AI, workflow rules).

Why do holes print small? Every FDM printer prints holes a little undersize. b123d-print adds your printer's hole compensation when it cuts holes, but it starts at 0 until you measure it with the hole coupon (Hole compensation). Don't also turn on the slicer's hole compensation (From part to print).

How accurate will it be? About as accurate as your printer and your measurements: ±0.2 mm is normal for FDM. The fits (clearance, press_fit, free_fit) are per printer and material; print the fit coupon once and store the values. Before a long print, test the fit with a quick fit-test template.

What if the part doesn't fit my printer's bed? Set your printer (Printer profiles) and the fit check fails when a part is too big; the assistant then proposes where to cut and how to join the pieces (Splitting parts).

How do I take it into my slicer? Open the .3mf in Bambu Studio, OrcaSlicer or PrusaSlicer; it is already in print orientation. Walls, supports, the Bambu "invalid config" message, several pieces and the print sheet: From part to print.

Can I edit the result in Fusion, Onshape or FreeCAD? Yes: open the .step, an exact solid without feature history. Changes there don't flow back to the script; the script stays the parametric source (Output formats).

Resin printers, CNC, laser cutting? No: the rules and checks are for FDM printing (overhangs, bridges, layer strength, bed edges). The STEP file is a normal solid you can take elsewhere, but nothing here checks it for those.

Where do the files go? Locally, into your workspace: parts/<name>/<name>.py (the script) and parts/<name>/out/ (STEP, STL, 3MF, PNG and section views). On a remote server they live on the server until download_part brings them home (Output formats, Which mode?).

The assistant says the part is done, but the files aren't in my project folder. On a remote server it must still bring them home: a remote export ends with a NEXT STEP saying so, but some agents skip it. Ask for download_part and an unzip into the project folder. If its shell has no network (Antigravity's agy with the terminal sandbox on), download_part(name, embed=True) returns the zip inside the result, and unzip works offline (Antigravity). Locally (stdio) the files are written straight into the project folder.

A big part times out. Each build has 240 s. The build's timing report names the slow piece; build one piece at a time with parts=, reuse a piece instead of building it twice, and use fewer points in curves (big and multi-piece projects). Raise --timeout only for a genuinely heavy part; MCP clients wait about 300 s for a result (server options).

The print came out wrong. See Print problems. Bring the photo back to the assistant: it can change the dimension with set_params and rebuild. To report a bug in b123d-print, open an issue with the part script and what went wrong.

Privacy, network and safety

Does it send my parts anywhere? No. Run locally, the server has no telemetry and opens no network connections of its own; the only network use is uvx downloading the package from PyPI. Your parts stay in your workspace. The assistant sees what you send it and what the tools return (text and renders), under its provider's terms. Remote modes (SSH, HTTP) send the parts to the server you set up, and nowhere else.

Does it work offline? The server and the 3D view do (the viewer ships inside the package and loads nothing). Your assistant usually doesn't, unless it is a local model.

The AI writes code that runs on my computer: is that safe? Part scripts run in a separate process with a time limit and, where the system has one, an OS sandbox: no network, writes only to the part's out/ folder (Sandbox).

It says "OS sandbox off", "no OS sandbox here" or "audit hook only". The scripts still run in a separate process with a time limit, but without the OS sandbox. On Linux install bubblewrap (apt install bubblewrap). On Windows there is no OS sandbox: use WSL2 with bubblewrap, or Docker, for scripts you don't trust. In the Docker image "sandbox off" is expected: the container is the sandbox. See Sandbox and Troubleshooting.

Windows

Does it work on Windows? Yes, with three things to know: there is no OS sandbox (above); an install can fail on a long path (turn on long paths, or set UV_CACHE_DIR to a short folder such as C:\uv); desktop apps may not find uvx (use the full path from where.exe uvx). Details in Troubleshooting.