swdraw

Turn a SolidWorks part into a finished, checked engineering drawing, and change parts from a spec with the same checks. Ask Claude, or run it yourself.

From part to drawing

1. Start with a part. Use your own, or generate one by script like the example flanged bush, which is built from a parameters file.

2. Describe the drawing in YAML. List the views, sections, details, dimensions, tolerances, GD&T, notes and title-block values. Each dimension also gets the value it should measure.

3. swdraw builds it. If a dimension measures wrong, it deletes it, regenerates the views and tries your alternative picks. If nothing works, it leaves the last attempt on the sheet and flags it.

4. Check the PDF and PNG. The result lists anything that failed.

5. Keep it current. Issue revisions, renumber parts without breaking references, tidy the working folder and update the drawing register.

Related MCP server: AutoCAD MCP Pro

Editing parts

Change an existing part from a short YAML spec: dimensions by full name (D1@Sketch1), global variables, equations, the configuration or a design-table row. swdraw rebuilds the part, then checks it:

• every changed dimension reads back what you set

• no feature has a rebuild error

• volume, mass, surface area and bounding box match your expect values

If anything fails, nothing is saved. With save: true it copies the original to <name>.bak.SLDPRT first. model inspect lists what a part has and writes a spec to start from. See references/modelling.md.

Planned: building parts from a spec (sketches, extrudes, revolves, cuts, holes, fillets, chamfers, patterns), with an expect after each feature.

Tools

Run it from the command line, from Python, or from Claude through the MCP server.

Requirements

• Windows with SolidWorks 2025 or later. Earlier versions probably work but aren't tested.

• uv (winget install astral-sh.uv). It installs Python and the dependencies on first run.

validate, model validate, tidy, register, number, setup and the tests don't need SolidWorks.

Install

git clone https://github.com/<you>/swdraw.git
cd swdraw
uv sync --extra mcp --extra preview

Then connect it to your AI client.

Claude Desktop extension. Get swdraw-<version>.mcpb from Releases (or build it with npx @anthropic-ai/mcpb pack . swdraw.mcpb) and open it in Claude Desktop. Set:

• Output folder: traces, results, PDFs and PNG previews

• Profiles folder (optional): your house_style.yaml, title_block.yaml, numbering.yaml

• Scripts folder (optional): the only folder swdraw_run_script can run scripts from

• Part backup folder (optional): where part edits copy the original before saving (default: next to the part)

One command from the clone. setup shows the change first. Add --apply to make it. Config files get a .bak backup.

uv run python -m swdraw setup claude-code --skill --output-dir ~/swdraw-output --apply
uv run python -m swdraw setup claude-desktop --output-dir ~/swdraw-output --apply   # not if you installed the .mcpb
uv run python -m swdraw setup codex --output-dir ~/swdraw-output --apply

--skill copies SKILL.md, references/ and subskills/ to ~/.claude/skills/swdraw so Claude Code knows the workflow. Restart the client after.

Other MCP clients. Run uv run --directory <clone> --extra mcp --extra preview python mcp_server/server.py over stdio. See mcp_server/README.md.

Profiles and template

Copy profiles/ somewhere and point SWDRAW_PROFILES (or the extension setting) at it. Anything missing falls back to the bundled files. See references/profiles.md.

For a matching A3 sheet format and drawing template, import templates/A3_SheetFormat.bas into a new SolidWorks macro, edit the constants at the top and run main. The constants cover company name, projection, confidentiality block, and linked properties or placeholders.

Quick start: the flanged bush

# 1. make the part and a drawing with front + isometric views
uv run python examples/flanged_bush/make_part.py --args "{\"out\": \"~/swdraw-example\"}"

# 2. check the spec (no SolidWorks needed)
uv run python -m swdraw validate --spec examples/flanged_bush/bush.drawing.yaml

# 3. with the new drawing active in SolidWorks, build it
uv run python -m swdraw build --spec examples/flanged_bush/bush.drawing.yaml

The result lists each step, what each dimension measured, which pick pair worked, and a check block with anything that failed. Look at the exported PNG too. The self-check proves the right edges got dimensioned, not that the layout is good.

To re-run part of the build, include clear or you'll get everything twice:

uv run python -m swdraw build --spec bush.drawing.yaml --steps clear,dimensions,gdt,leader_notes,notes,fonts,export

Widen the flange to 42 mm and check the new volume (add --save to keep it):

uv run python -m swdraw model edit --spec examples/flanged_bush/bush.edit.yaml

Issue revision B:

uv run python -m swdraw revise --args "{\"rev\": \"B\", \"desc\": \"BORE TOLERANCE CHANGED\", \"drawn\": \"AD\", \"approved\": \"CC\", \"approver_name\": \"C. CHECKER\"}"

Writing a spec

Anything on geometry (pick, text_at, section line, detail centre) is in model mm. Positions on the sheet (at) are sheet mm from the bottom-left corner. Every dimension has an expect value.

Start from the spec template. The flanged bush spec is a full example, and references/drawing-spec.md covers every key.

Your own scripts

The helpers are small, so one-off scripts are easy:

import swdraw as sw

dwg = sw.active()
names = [sw.call(v, "GetName2") for v in sw.views(dwg)]
sw.result({"views": names, "args": sw.args()})

Run it with uv run python -m swdraw run my_script.py --args "{...}", or from Claude with swdraw_run_script if it's in your scripts folder.

When something goes wrong

Check references/troubleshooting.md, then references/api-gotchas.md. Most surprises are the SolidWorks API.

uv run python -m swdraw diag views          # names, origins, scales, whether model +Y is up
uv run python -m swdraw diag annotations    # incl. dangling centre marks

Every run logs to swdraw.log in the output folder.

Repository layout

Tests

uv run --extra dev python -m pytest

The tests run against a fake SolidWorks API (tests/fakes.py). They cover:

• selection and dimension self-checking, including the fallbacks and tolerances

• title-block matching, revisions and notes

• tidy, the register (including formula results when LibreOffice is installed) and numbering

• spec validation and the example's geometry

• part edits: globals, equations, dimensions, configurations, design tables, read-back, rebuild errors, expect, revert, backup and save

CONTRIBUTING.md lists what they can't cover. CI runs them on Windows and Linux on every push.

Licence

MIT. See LICENSE.