# CADCLAW

CADCLAW is an automated check suite for STEP-based CAD assemblies — "pytest for
mechanical design", in spirit. It is a Python package (PyPI: `cadclaw`) that
validates STEP files and BOM JSON through a chain of geometric, kinematic,
tolerance, and text-audit gates, and reports findings with severity, evidence,
and a confidence budget rather than just pass/fail. CADCLAW also ships an MCP
server so MCP-compatible assistants (Claude Desktop, Cursor, etc.) can call its
checks directly.

This file is a high-density plain-text summary intended for LLM and
agent-driven retrieval. The canonical site is https://cadclaw.io.

## Identity

- Name: CADCLAW
- Publisher: Sunnyday Technologies (https://sunn3d.com)
- Authored and maintained by: Nick Sonnentag, Sunnyday Technologies
- Contact: info@sunn3d.com
- Repository: https://github.com/sunnyday-technologies/CADCLAW
- License: MIT
- PyPI: https://pypi.org/project/cadclaw/
- DOI: 10.5281/zenodo.19647391
- Current version: 0.7.1
- Site: https://cadclaw.io
- Status: open source, actively maintained

## What CADCLAW does

CADCLAW validates STEP CAD assemblies and their BOMs through automated gates.
Configure once in `cadclaw.yaml`, run from the `cadclaw` console script.

Validation domains (each maps to a module under `cadharness/`):

- Inventory (`cadharness/inventory.py`) — label parts by bbox signature,
  count against expected, per-region (axis-aligned) constraints.
- Interference (`cadharness/interference.py`) — pairwise solid-solid overlap
  via OCC `BRepAlgoAPI_Common`, reports overlap volume in mm^3.
- Adjacency (`cadharness/adjacency.py`) — parts of type A must have a part
  of type B within N mm. Catches misplaced/scattered components.
- Dimensional (`cadharness/dimensional.py`) — checks part dimensions against
  expected ranges; catches wrong thickness, swapped args, scaling errors.
- Kinematics (`cadharness/kinematics.py`) — beam deflection (Euler-Bernoulli),
  motor torque budgets, belt tension, GT2 tooth-skip resistance.
- Tolerance (`cadharness/tolerance.py`) — worst-case, RSS, and Monte Carlo
  tolerance stacking with Cpk and per-dimension variance contribution.
- Parity (`cadharness/parity.py`) — STEP-vs-STEP comparison; flags the
  Fusion visibility-toggle bug.
- BOM audit (`cadharness/bom_audit.py`) — BOM JSON ↔ CAD assembly: qty,
  mfg_type, unit, required/forbidden text terms, CAD-side count drift.
- Disassembly (`cadharness/disassembly.py`) — sequenced part removal,
  radial exploded views, animation frame export.
- Render (`cadharness/render.py`) — STEP → PNG → animated GIF via
  offscreen VTK.

Honesty toolchain (text-side validation, also surfaced via the CLI):

- `cadclaw doctor` — environment diagnostic.
- `cadclaw publish-audit` (`cadharness/publish_audit.py`) — scans the working
  tree for private data before commit.
- `cadclaw claim-audit` (`cadharness/claim_audit.py`) — flags overclaims and
  untagged numeric assertions in README and BOM notes.

CADCLAW does NOT prove that a native CAD model has no hidden parts, that a
physical build matches CAD, that a vendor part is in stock, that a printed
part is strong enough for production, or that an AI-generated CAD change is
correct. Each report includes a confidence budget per gate.

## Install

```
pip install cadclaw
cadclaw doctor   # verify environment
```

Requirements: Python 3.10+, CadQuery 2.7+ (provides OCC/STEP support),
pyyaml 6+, pydantic 2.5+. No commercial CAD software is required for
CADCLAW's own checks. Cross-platform (Windows, macOS, Linux).

## CLI

```
cadclaw doctor
cadclaw harness       --rules cadclaw.yaml
cadclaw bom-audit     --rules cadclaw.yaml
cadclaw publish-audit --rules cadclaw.yaml
cadclaw claim-audit   --rules cadclaw.yaml --report-format md -o report.md
cadclaw inspect       ...                 # diagnostic queries
```

Exit codes: 0 pass, 1 fail, 2 warn-only (no fails), 3 internal error.

## MCP server

The repository also publishes `cadclaw_mcp/` — a Model Context Protocol
server that exposes the same harness modules as MCP tools, so an
MCP-compatible assistant can drive CADCLAW directly. The MCP server runs
LOCALLY after `pip install cadclaw`; it is NOT a hosted endpoint. It only
exposes CADCLAW's own checks — it does not give an assistant access to your
CAD application or to anything outside what `cadclaw` itself can do.

Entry point: `python -m cadclaw_mcp` (see `cadclaw_mcp/server.py`).

## Companion projects

- M3-CRETE — open-source concrete 3D printer; CADCLAW's first production
  deployment. https://m3-crete.com · https://github.com/sunnyday-technologies/M3-CRETE
- CEMFORGE — CAM / slicer surface for the M3-CRETE family. https://cemforge.ai
- Open3DCP — open 3D-construction-printing standards effort.
  https://open3dcp.org

## Machine-readable surface

- Site root (HTML, with embedded SoftwareApplication JSON-LD):
  https://cadclaw.io/
- Sitemap: https://cadclaw.io/sitemap.xml
- This file: https://cadclaw.io/llms.txt
- MCP manifest (experimental, static-site read-only resources):
  https://cadclaw.io/.well-known/mcp-manifest.json
- robots.txt (explicit opt-in for ClaudeBot, GPTBot, Google-Extended,
  PerplexityBot): https://cadclaw.io/robots.txt
- README, source, issues: https://github.com/sunnyday-technologies/CADCLAW
- Package: https://pypi.org/project/cadclaw/

## Citation

```
Sonnentag, N. (2026). CADCLAW: Automated validation framework for
STEP-based CAD assemblies. Sunnyday Technologies.
https://github.com/sunnyday-technologies/CADCLAW
DOI: 10.5281/zenodo.19647391
```