WoodShop

Voice-driven conversational 3D woodworking & furniture modeler.

Talk to it like the Star Trek holodeck and watch furniture build itself:

"Place a 6 foot 2x4, sand it, then attach a 2 foot 2x4 at 90 degrees, 10 inches from the end."

"Build a coffee table: a four foot by two foot frame from 2x4s, with four legs 18 inches tall standing at the corners."

Each board is real dimensional lumber (a 2x4 is modeled at its true 1.5″ × 3.5″), so the result is buildable — export to STEP (CAD/CNC) or STL (3D print), and get a cut list with board-feet and a shopping estimate.

How it works

WoodShop reuses the existing CmdForge tool ecosystem for everything that isn't woodworking-specific, so no wheels are reinvented:

woodshop-talk ── the conversational loop
   │  dictate ............. speech → text          (CmdForge tool)
   │  pa-load-tools ....... wood-* → Claude schemas (CmdForge tool)
   │  claude -p ........... interpret → tool calls  (provider)
   │  pa-execute-tool ..... dispatch each wood-*     (CmdForge tool)
   │  read-aloud .......... speak confirmation       (CmdForge tool)
   ▼
scene.json  ← single source of truth (parts, joints, selection, undo)
   ▲                                   │ writes
   │ reads/mutates                     ▼
wood-* CmdForge tools          woodshop-view
(place/join/stand/move/...)    live pyvista 3D, watches scene.json

The wood-* tools are thin wrappers over the woodshop CLI, so the modeling logic lives in one place and the tools double as the LLM's documented command vocabulary.

Installation

python -m venv .venv && source .venv/bin/activate
pip install -e ".[gui,dev]"        # 'gui' pulls build123d + pyvista + PySide6 + pyvistaqt
python scripts/gen_wood_tools.py   # register the wood-* CmdForge tools

Usage

The studio (recommended)

woodshop            # launches the unified desktop app

One window with the 3D viewport (click a board to select it), a parts panel (list + selected-part inspector + quick-action buttons), and a command bar at the bottom where you type or push-to-talk (🎤). Mouse, keyboard, and voice all drive the same scene and the same visible selection, so "delete that" / the Delete button / saying "delete the front-left leg" are interchangeable. Menus cover New/Open/Save projects, Export STL/STEP, Save Image, Undo/Redo, camera views, and Build templates.

Standalone tools (headless / scripting)

woodshop-view &                    # just the live 3D window (watches the scene)
woodshop-talk                      # just the voice/text loop; --voice to speak
woodshop-talk --once "build a workbench top from five 2x6 boards 6 feet long"

Or drive it directly from the CLI:

woodshop place 2x4 "6 ft"          # place a board
woodshop stand                     # stand it up (a leg)
woodshop join p2 --to p1 --angle 90 --offset "10 in"
woodshop rename "front-left leg"
woodshop cutlist                   # bill of materials
woodshop export table.step         # STEP / STL export
woodshop save "coffee table"       # named projects
woodshop open "coffee table"

Run woodshop --help for the full command list (place, join, stand, lay, rotate, move, trim, copy, rename, sand, delete, undo, clear, status, cutlist, export, save, open, projects).

The active scene lives at $WOODSHOP_SCENE or ~/.local/share/woodshop/scene.json; named projects in ~/.local/share/woodshop/projects/.

Development

pytest                              # 41 tests

Key modules:

Module	Role
scene.py	Part/Joint/Scene model, operations, undo, persistence
lumber.py	nominal → actual dimensional lumber table
units.py	parse "6 ft" / "3 ft 6 in" / "-2 ft" → inches
cli.py	the woodshop command
geometry.py	build123d solids + STL/STEP export
cutlist.py	cut list, board-feet, shopping estimate
viewer.py	live pyvista 3D viewport (woodshop-view)
driver.py	conversational loop (woodshop-talk)
scripts/gen_wood_tools.py	(re)generate the wood-* CmdForge tools

Known limitations

• Joins are butt joints (B's end sits flush against A's face, not its centerline); joinery cuts (mortise/tenon, lap, pocket holes) aren't modeled yet.
• Command interpretation latency is ~7–13s per utterance (one claude -p call).

License

MIT