onshape-mcp

# onshape-mcp I wanted a way to drive Onshape from a language model the way I drive it with my hands. Click a face. Draw a sketch. Extrude. Chamfer. The existing Onshape MCP wraps the REST API and FeatureScript, which works but feels like coding, not modeling. This server aims at the other 70% of CAD work that lives in the viewport. It uses a vision-capable LLM (Gemini web, no API key, my Plus cookies) as eyes, Playwright as hands, and an MCP surface so any MCP-aware client can talk to it. Every action is journaled, so I can undo, replay, or branch. ## Status - **M0** done: scaffold, smoke tests, public repo, safety rails. - **M1** done: driver primitives, Onshape tool datasheet, 43 MCP tools across sketching, constraints, dimensions, features and feature-tree editing, plus the closed-loop `act(goal)` agent and a deterministic fast path that skips the LLM entirely for parseable goals. - **M2** next: pattern, mirror_body, assembly.mate, sketch.constrain flyout, journal-replay undo, a real perceptual-diff for the stuck detector. ## A note on sizes Sizes come from Onshape's dimension solver, not from pixel measurements. Tools draw a rough shape at a scale derived from the live canvas, then drive the true millimetre value in — so a rectangle or circle is exactly the size you asked for regardless of zoom. All dimension arguments are millimetres unless you write a unit (`"10 cm"`, `"2 in"`); bare numbers are never reinterpreted. Polygon radius is currently drawn-to-scale but not solver-driven; see the `ponytail:` note in `ui_actions.py`. ## What it looks like from the client side Two ways to drive it: **Direct tools.** I call individual MCP tools when I want fine control: `screenshot`, `describe_view`, `viewport_size`, `journal_tail`, `tool_datasheet`, `open_doc`, and the per-tool `onshape_*` ones. Useful when I want to see each step and steer. **Closed-loop `act`.** I just say what I want: `act(goal="draw a 50x30mm rectangle on the top plane and extrud…