cadclaw 0.10.0

CADCLAW: automated assembly and validation for STEP CAD. Places authored parts by declared constraints, then runs geometric, structural, tolerance, and BOM-vs-CAD gates. Includes an MCP server.

CADCLAW

Build STEP CAD assemblies from authored parts, then check them automatically.

Generated end-to-end with render_radial_explode_gif("M3-2_Assembly.step", "out.gif") — parts explode radially from the centroid, then camera orbits 360°. 99+ parts, no manual animation work.

CADCLAW does two things. It assembles: a declarative spec places your authored STEP parts by connector frames and datum chains, compiles the assembly with CadQuery, and emits a design inventory, a model-derived BOM, review renders, and step-by-step build sequences. It verifies: automated gates (inventory, interference, adjacency, dimensional, orientation, floating-part, color/material, structural, tolerance stacking, parity) plus a BOM-vs-CAD audit and an honesty toolchain (doctor, publish-audit, claim-audit).

CADCLAW places parts you authored in real CAD. It does not generate geometry: no parametric plates, no bolt-circle helpers, no hole patterns. You draw the parts, CADCLAW seats them against each other and checks the result.

Like pytest for mechanical design, in spirit. Real CAD has analog characteristics pytest doesn't have (a part isn't binary present/absent, it can be slightly the wrong size, slightly clipping, slightly misplaced), so CADCLAW reports findings with severity, evidence, and a confidence budget rather than just pass/fail.

CADCLAW is also the open verification engine behind MARB — the Mechanical Assembly Readiness Benchmark — whose graders import CADCLAW's gates to score how well AI assembles a complete machine in CAD. See the MARB repository.

The Problem

CAD assemblies break silently. Parts clip into each other, BOMs drift from geometry, motor mounts end up 600mm from the motor. Engineers catch these errors by eye — if they catch them at all. CADCLAW automates the geometric checks. It does not replace engineering judgment, structural certification, or physical-build validation.

What CADCLAW Does

1. Assemble

An assembly spec (assembly_spec.v0.1) declares the parts, where they come from, and how they seat against each other. cadclaw assemble resolves that spec and compiles it into a STEP assembly with CadQuery.

Command	What it does
assemble validate-spec	Validate a spec before compiling. Unknown keys fail; incomplete work is declared explicitly as not_built_yet.
assemble build	Resolve authored STEP sources and compile the assembly. --dry-run resolves paths without touching geometry.
assemble check-round	Build, inventory-check, render review views, and report one assembly round. The main iteration loop.
assemble inspect-component	Inspect one authored STEP component: bbox signature, part count, isolated review renders.
assemble render-views	Render the review_views a spec declares (iso, hero, front, side, top, and more).
assemble render-sequence	Export partial assembly STEPs, per-step review views, a BOM CSV, and an optional rotating GIF.

Parts are placed by constraint, not by hand-typed coordinates. An instance declares place_relative_to: seat this connector frame against that parent frame, offset along an axis. The resolver walks the datum chain in topological order and solves each transform, reporting cycles and missing frames as findings. Absolute transforms still work, so migration is incremental. lock: axis solves only the handoff axis for parts that span the other two, like a gantry.

Connector frames (extrusion ends, mount faces, rail slots, wheel contacts, shaft axes, belt planes) are recorded per component in connector metadata. This is descriptive data about parts you authored. It does not generate contextual geometry.

Assembly outputs are non-authoritative by design: protected_paths stops a build from overwriting your real CAD exports.

Docs: The assembly spec is the field-by-field reference. examples/relative_placement/ is a small runnable example of constraint placement: three parts, one datum, both lock modes, with the solved coordinates asserted in tests.

cadclaw assemble check-round examples/relative_placement/gantry.yaml

2. Verify

CADCLAW validates STEP assemblies + BOM JSON through a chain of automated gates:

Gate	What it catches
Inventory	Missing/extra parts. Labels by bbox signature, counts against expected. Per-region (axis-aligned) constraints supported.
Interference	Solid-solid overlaps. BRep boolean intersection, not just bbox.
Adjacency	Parts that should be near each other but aren't (motor 600mm from mount).
Dimensional	Wrong thickness, swapped box() args, impossible dimensions.
Orientation	Rotated/mis-faced parts where a label declares an expected face plane.
Floating	Parts isolated from configured structural labels beyond a max gap.
Color/material	STEP AP242 color metadata against expected label colors.
Structural	Beam deflection, motor torque budget, belt tension. Static load math, not motion-clearance or full-travel sweeps.
Tolerance	Worst-case, RSS, Monte Carlo tolerance stacking with Cpk and variance decomposition.
Parity	STEP-vs-STEP comparison; flags hidden/suppressed-part export drift.
BOM audit	BOM JSON ↔ CAD assembly: qty, mfg_type, required/forbidden text terms, CAD-side count.
Disassembly	Sequenced part removal, radial exploded views, animation frame export.
Render	STEP → PNG → animated GIF via offscreen VTK.

The CLI harness runs the checks declared in cadclaw.yaml; geometry checks share the STEP export when possible, while parity, render, disassembly, tolerance, and audits are also available as focused commands/APIs. Every report includes a confidence budget that lists what was checked, what was not, and what assumptions were made.

CADCLAW also includes an MCP Server: 23 tools covering both halves, so an MCP-compatible assistant can drive CADCLAW directly. The six assemble_* tools build and inspect assemblies; the rest run the checks. The MCP server does not give the assistant access to your CAD application or to anything outside what cadclaw itself can do.

The render-producing assembly tools return their PNGs as inline images, so the assistant can look at what it just built instead of trusting a path string. Every render is also written to disk, giving the human a per-step traceability artifact of what changed and when.

The full loop is: prompt, edit the assembly spec, assemble check-round, inspect the report and the review renders, repeat.

What CADCLAW Does NOT Prove

CADCLAW checks the geometry of a STEP file, the JSON of a BOM, and the text of your README against rules you write. It does not prove:

• That the native CAD model has no hidden or suppressed parts. CADCLAW reads the STEP export, which can silently drop invisible parts.
• That the physical build matches the CAD. CAD passing CADCLAW says nothing about whether the parts on your bench match the file.
• That a vendor part is in stock, available, or the price you assumed.
• That a printed part is strong enough for production use. CADCLAW's kinematics gates do bare-beam math; they don't simulate printed-PLA fatigue, layer adhesion, or thermal creep.
• That a structural claim is physically certified, unless you've attached measurement data with an evidence tag.
• That an AI-generated CAD change is correct without passing the gates. CADCLAW is the check; not passing it doesn't make a change correct, only "passed the gates we have."

Each report includes a confidence budget per gate: checked, not_checked, assumptions. Read it.

Honesty toolchain

• cadclaw doctor — environment diagnostic. Run this first.
• cadclaw publish-audit — scans the working tree for private data before you commit.
• cadclaw claim-audit — text linter that flags overclaims and untagged numeric assertions in your README and BOM notes.

These three tools exist because the truthfulness of CADCLAW's reports is only as good as the truthfulness of the docs and BOM that surround them.

Using CADCLAW with an AI assistant

If an AI assistant is editing your CAD code, point it at AGENTS.md. The short version: place authored parts; do not generate them. CADCLAW verifies geometry the user authored in native CAD tools; only genuinely parametric stock (extrusion bars, V-wheels) should ever be generated by the assistant. AGENTS.md exists because field tests showed that AI-generated plates and motor mounts can ship with hole patterns that do not align with their assemblies.

For diagnostic queries (signature histogram, "what is this part", "what overlaps with X"), use cadclaw inspect rather than writing throwaway probe scripts.

Quick Start

pip install cadclaw
# cadquery, vtk, Pillow, pyyaml, pydantic are pulled in automatically.
# For editable dev installs:
#   git clone https://github.com/sunnyday-technologies/CADCLAW.git
#   cd CADCLAW && pip install -e .

cadclaw doctor                          # verify your environment first

Programmatic API

from cadclaw.harness import Harness
from cadclaw.adjacency import AdjacencyRule

h = Harness("my_assembly.step")

h.add_inventory(
    labels={(40.0, 80.0, 1000.0): 'beam', (56.4, 56.4, 76.6): 'motor'},
    expected={'beam': 4, 'motor': 2, 'belt': 3}
)

h.add_interference(skip_labels={'belt', 'wheel'})

h.add_adjacency(rules=[
    AdjacencyRule('motor', 'bracket', max_distance=50)
])

report = h.run()
print(report)
# CAD HARNESS REPORT — PASSED
#   Parts: 42
#   Time:  3200ms
#
#   [PASS] inventory (120ms)
#   [PASS] interference (2800ms)
#   [PASS] adjacency (15ms)

CLI workflow

Configure once in cadclaw.yaml — labels, expected inventory, regions, BOM rules, claim-audit terms, publish-audit globs — then drive everything from the cadclaw console script:

cadclaw doctor                                    # 1. verify the environment
python examples/init_rules.py --step my.step      # 2. scaffold cadclaw.yaml
                              --bom bom.json
cadclaw harness --rules cadclaw.yaml              # 3. run configured YAML-backed checks
cadclaw bom-audit --rules cadclaw.yaml            # or run a single gate
cadclaw publish-audit --rules cadclaw.yaml        # before `git push`
cadclaw claim-audit --rules cadclaw.yaml --report-format md -o report.md

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

BOM-vs-CAD audit (the v0.6 headline)

# cadclaw.yaml fragment
bom_audit:
  bom_path: bom/data.json
  rules:
    - id: 5
      expected_qty: 12
      expected_label: connector_bar
      forbidden_terms: ["maximum rigidity", "primary stiffness"]
    - id: 65
      expected_qty: 3
      expected_unit: "bars (1.0m each)"
      expected_mfg_type: buy
      required_terms: ["1m", "friction-fit"]
      forbidden_terms: ["JB Weld", "West System", "custom 2m cut"]

The audit catches:

• BOM qty / mfg_type / unit mismatches
• Required-term-missing / forbidden-term-present in name + description + notes
• CAD-side count drift (CAD has 16 connectors, BOM expects 12)
• BOM items with no CAD geometry (suppressed for mfg_type: consumable / electronic / fastener)
• CAD parts with no covering BOM rule

Private BOM fields (vendors, sku, unit_cost, anything starting with _) are dropped at the serializer level and never appear in any report.

How It Works

Every solid in a STEP file has a bounding box. The sorted dimensions (dx, dy, dz) rounded to 0.1mm form a signature — a fingerprint that identifies part types without needing part names or metadata.

(40.0, 80.0, 1000.0) → "beam"     # 4080 C-beam extrusion
(56.4, 56.4, 76.6)   → "motor"    # NEMA23 stepper
(4.0, 80.0, 96.0)    → "mount"    # motor mount plate

This works because mechanical parts have characteristic dimensions. A NEMA23 is always 56.4mm square. A 4080 extrusion is always 40x80mm. The harness exploits this invariant to label, count, and validate without parsing STEP metadata.

Author

CADCLAW is authored and maintained by Sunnyday Technologies, led by Nicholas Sonnentag (ORCID 0009-0002-1897-384X). Development uses Sunnyday Technologies' LLM-assisted engineering practice; design decisions, engineering judgment, test fixtures, and direction are owned by the Sunnyday Technologies team.

Contact: [e-mail]

Citation

If you use CADCLAW in published research or derivative work, please cite:

Sonnentag, N. (2026). CADCLAW: Automated validation framework for
STEP-based CAD assemblies. Sunnyday Technologies.
https://github.com/sunnyday-technologies/CADCLAW
DOI: 10.5281/zenodo.19647390

A CITATION.cff file is included for automated citation tooling.

Origin Story

CADCLAW was developed alongside the M3-CRETE open-source concrete 3D printer project — built out of a practical need to properly position and validate components during assembly of a large, part-dense machine, using Sunnyday Technologies' LLM-assisted engineering practice. The harness:

• Caught 53 solid-solid interferences in a single run
• Reduced STEP file size from 70MB to 13MB by identifying geometry bloat
• Checked 150+ assembly changes across 15 design sessions without visual inspection [analysis]
• Prevented 3 regressions that would have shipped broken geometry to builders

See examples/m3_crete/ for the reference implementation.

Modules

cadclaw.assembly_spec

The declarative contract a human or LLM edits before compilation. Strict pydantic schema: unknown keys fail validation, generated outputs cannot overwrite protected CAD exports, and incomplete work is represented explicitly as not_built_yet. Defines Instance, Transform, RelativePlacement, and ReviewView.

cadclaw.assembly_compiler

Resolves a spec into geometry. resolve_relative_placements() walks the datum chain topologically and solves constraint-placed transforms; run_assembly_build() compiles the STEP; run_assembly_check_round() builds, checks, and renders one iteration; run_assembly_sequence() exports the step-by-step build. Also writes the design inventory and BOM CSV.

cadclaw.connector_metadata

Local coordinate frames per authored component: extrusion ends, mount faces, rail slots, wheel contacts, shaft axes, belt planes. The bridge between an authored STEP asset and reliable constraint placement. Descriptive only; it does not author geometry.

cadclaw.component_manifest

Observational index of an authored STEP library: where assets live, their bbox signatures, and which entries still lack BOM or connector metadata.

cadclaw.inventory

Label parts by bbox signature, count them, compare to expected inventory.

cadclaw.interference

Pairwise solid-solid overlap using OCC BRepAlgoAPI_Common. Bbox pre-filter for performance. Reports overlap volume in mm^3.

cadclaw.adjacency

Validate that parts of type A have a part of type B within N mm. Catches misplaced/scattered components.

cadclaw.dimensional

Check part dimensions against expected ranges. Catches wrong thickness, swapped args, scaling errors.

cadclaw.kinematics

Structural load math from assembly parameters: beam deflection (Euler-Bernoulli), motor torque budget, and belt tension against breaking/working limits. Static analysis only; it does not sweep range of motion or check clearance through travel.

cadclaw.tolerance

Tolerance stack analysis: define dimension chains, compute worst-case / RSS / Monte Carlo accumulation, report Cpk process capability and per-dimension variance contribution. Identifies which dimension dominates the stack.

cadclaw.disassembly

Disassembly sequence generation: auto-orders parts by type priority and distance from centroid, computes radial explosion vectors, exports individual STEP frames for animation or a single exploded-view STEP.

cadclaw.render

Offscreen VTK rendering of STEP files to PNG, plus GIF stitching. make_disassembly_gif(step, gif) is one call — generates the disassembly frames, rasterizes them, and writes an animated GIF.

cadclaw.harness

The runner. Chains gates, loads parts once, reports pass/fail with timing.

cadclaw_mcp/

MCP Server exposing CADCLAW's assembly and validation tools (23) to MCP-compatible hosts. The six assemble_* tools cover spec validation, compilation, the check round, component inspection, review rendering, and sequence export; the remainder run the checks and audits. Render-producing tools return PNGs as inline image content so the assistant can visually verify each round. No code generation needed — MCP is an open protocol, so any compliant client can drive the harness.

CI/CD Integration

# .github/workflows/cad-check.yml
- name: Validate assembly
  run: |
    pip install cadclaw
    python check.py assembly.step

Exit code 0 = passed. Exit code 1 = failed. Works in any CI system.

Who This Is For

• Open-source hardware projects — catch assembly errors before builders hit them
• CadQuery/FreeCAD users — the testing layer the ecosystem is missing
• Small manufacturing teams — automated QA between design and procurement
• AI-assisted CAD workflows — validate that AI-generated changes don't break the assembly

Running Tests

git clone https://github.com/sunnyday-technologies/CADCLAW.git
cd CADCLAW
python -m venv .venv
.\.venv\Scripts\python -m pip install --upgrade pip
.\.venv\Scripts\python -m pip install -e .

# Generate test fixture STEP assemblies (L1-L3, good + bad variants)
.\.venv\Scripts\python tests/generate_fixtures.py

# Run the full test suite
.\.venv\Scripts\python -m unittest discover tests

The test fixtures are generated from CadQuery — no external downloads needed. Three tiers of increasing complexity:

Level	Parts	Tests
L1: Bracket assembly	5	Inventory, interference
L2: Motor mount	10	Inventory, adjacency
L3: Gantry corner	18	Full 4-gate harness

Each level has a "good" variant (should pass) and "bad" variant (deliberate errors for the harness to catch: clipping, missing parts, scattered motors).

The suite also exercises tolerance stacking math against hand-calculated answers, the full disassembly pipeline, the MCP server over real JSON-RPC, and end-to-end GIF rendering.

Requirements

• Python 3.10+; Python 3.11 is the current CADCLAW development runtime
• CadQuery 2.7+ (provides OCC/STEP support)
• VTK 9.3+ for rendering, plus Pillow 10+, pyyaml 6+, and pydantic 2.5+ (pulled in automatically)
• No commercial CAD software needed for CADCLAW's own checks. Validation that depends on the native CAD application — feature-tree review, native-format parametric checks — is outside CADCLAW's scope.

Run cadclaw doctor after install to verify your environment.

License

MIT License. Copyright (c) 2026 Sunnyday Technologies.

Built during the M3-CRETE project — an open-source concrete 3D printer where CADCLAW caught 53 interferences, reduced STEP file size from 70 MB to 13 MB, and ran 150+ assembly checks across a human-AI design collaboration [analysis].