Link Search Menu Expand Document
OCCTMCP

OCCTMCP Cookbook

Task-oriented, example-rich recipes for driving OCCTMCP from an MCP client (an LLM). One page per area: a short bit of prose, then the actual tool calls — each shown as the JSON arguments you send and an example JSON result you get back — chained into a real workflow. Figures for geometry-building recipes are rendered by OCCTMCP’s own render_preview and committed under images/.

This is the usage counterpart to the per-tool Tool Reference; recipes link to the reference rather than restating every parameter.

Conventions

  • Show real tool calls. Each step is a ```json block of the arguments object for one tool, optionally followed by a ```json example result. Use only real parameters (see the Tool Reference / the tool’s schema) — never invent fields.
  • Typed tools first. Reach for execute_script only when authoring geometry the typed tools don’t cover; say so when you do, and keep scripts to the script template.
  • One canonical place per topic. Recipes hold workflow; per-tool detail lives in the reference, architecture in Architecture. Link, don’t duplicate.
  • Note the server. When a recipe uses a Swift-only tool (selection / remap / annotations / reconstruct / graph_select / pick_surface_point), say it needs the Swift occtmcp-server.

Figures & interactive 3D

OCCTMCP renders its own figures from the same tool calls the page shows, so code and figure never drift. Build the scene with execute_script (or the typed construction tools), then:

  • PNG via render_preview → committed under images/ (also the loading poster below).
  • Interactive GLB via export_scene (format: "glb") → committed under models/, embedded with Google’s <model-viewer> web component (orbit / zoom / auto-rotate), using the PNG as the poster until the model loads.
<script type="module" src="https://cdn.jsdelivr.net/npm/@google/model-viewer/dist/model-viewer.min.js"></script>
<model-viewer src="models/<name>.glb" poster="images/<name>.png" alt="…"
  camera-controls auto-rotate environment-image="neutral" exposure="1.1" shadow-intensity="1"
  style="width:100%;max-width:480px;height:360px;background:#eef1f5;border-radius:6px"></model-viewer>

Pages

  • Authoring with execute_script — the script template, building geometry against the full OCCTSwift API, get_script, and the scene/manifest model.
  • Scene & appearance — read the scene, recolor / rename / remove bodies, and diff versions.
  • Construction — features (drill / fillet / chamfer / extrude / revolve / thread), transforms, booleans, and mirror / pattern.
  • Inspection — validate geometry, compute metrics (volume / area / bbox / boundingBoxOptimal), query topology, and recognize features.
  • Measurement & verification — minimum distance vs. surface deviation (Hausdorff) for certifying a reconstruction, wall thickness, and clearance.
  • Selection & remap — pick faces / edges / vertices to stable selectionIds and carry them across mutations (Swift only).
  • Annotations & preview — dimensions, scene primitives, bounding boxes, diff overlays, and PNG previews / pixel picking (Swift only).
  • Import, export & assemblies — load BREP / STEP / IGES (incl. allowInvalid), export the scene, and walk / edit XCAF assemblies.
  • Meshing & drawings — tessellate, decimate, and produce multi-view technical drawings.
  • Healing — repair imported / non-watertight geometry and read before/after stats.
  • Topology graph — validate / compact / dedup the B-rep graph, export ML-friendly JSON, and local adjacency / selection.
  • Reconstruction graph — read/annotate/persist the attributed reconstruction graph (Swift only).

Table of contents