by baney75
A skill and CLI that lets small local models like LFM3-1.7B make 3D printable parts, from words or a picture, even through context compaction.
# Add to your Claude Code skills
git clone https://github.com/baney75/smoothstoneGuides for using ai agents skills like smoothstone.
See how smoothstone compares with popular alternatives.
smoothstone is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by baney75. A skill and CLI that lets small local models like LFM3-1.7B make 3D printable parts, from words or a picture, even through context compaction. It has 58 GitHub stars.
smoothstone's catalog security scan is still queued. You can run an instant dependency and prompt-injection check now with the "Scan for vulnerabilities" button above.
Clone the repository with "git clone https://github.com/baney75/smoothstone" and add it to your Claude Code skills directory (see the Installation section above).
smoothstone is primarily written in Python. It is open-source under baney75 on GitHub, so you can review or fork the full source.
Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh smoothstone against similar tools.
No comments yet. Be the first to share your thoughts!
Available after the catalog security scan passes.
⚠️ Third-Party Software Notice
This skill is third-party open-source software developed and hosted independently on GitHub. SkillsLLM is an informational directory and does not control or maintain the underlying repository.
Any security checks, ratings, or warnings displayed by SkillsLLM are automated and limited in scope. They do not constitute a security certification or guarantee that the software is safe, error-free, or free from malicious code, vulnerabilities, compromised dependencies, or prompt-injection risks.
Review the source code, permissions, dependencies, and configuration before installing or running any third-party skill. Use is at your own risk. To the maximum extent permitted by applicable law, SkillsLLM is not liable for losses arising from third-party software.
The deep catalog scan for this skill is still queued. Run an instant dependency check now instead.

I wanted a small model running on my own computer to make 3D printable parts that actually look good, so I tested LFM3-1.7B writing OpenSCAD on its own and it failed every single time. Even with the best skill I could write it made a working fan guard only 2 times out of 5. smoothstone takes the hard part away from the small model, as the model never writes geometry at all. It picks a tested recipe and fills in the sizes from the request or a picture, and the stone tool builds the part, checks that it will print and keeps every step on disk, so the model can lose its whole context and still finish the part.

LFM3-1.7B made every part above on my own CPU through smoothstone, and each one passed every print check. I did not want to fool myself with requests I had tuned for, so a fresh agent that never saw the code wrote new requests from the parameter list alone and I sealed them before the run. Each request ran five times. On the last blind set the model finished the right part with the right sizes 39 out of 40 times, and 38 out of 40 when I wiped its memory in the middle of the job. Writing the 3D code itself it went 0 out of 15. Every set, every failed run and every fix is in LOG.md, and the sets are in evals.
Every stone message ends with one line that says what to run next, and stone status prints that line again along with the request, the recipe, the sizes so far and the last build. A model that forgets everything runs stone status and keeps going.
$ stone new "A guard for a 120 mm PC fan, 3 mm thick."
NEXT: stone recipes
$ stone recipes
fan_guard Fan guard (default size 120 x 120 x 3 mm)
NEXT: stone show fan_guard
$ stone show fan_guard
From the request: size=120 thickness=3
NEXT: stone set fan_guard size=120 thickness=3
$ stone set fan_guard size=120 thickness=3
Expected size 120 x 120 x 3 mm.
NEXT: stone build
$ stone build
Size: 120.0 x 120.0 x 3.0 mm
PASS watertight: closed surface
PASS bodies: 1 bodies (max 1)
PASS first_layer: 6344.7 mm2 on the bed (need 200.0 mm2)
PASS min_wall: thinnest wall 1.98 mm (need 0.8 mm)
PASS overhang: 0.0% of the surface hangs past 45 deg (max 5%)
PASS dims: size matches the recipe
Part ready: stone-job/fan_guard.stl
NEXT: none, the part is ready
That is shortened: each command prints a few more lines, and the build prints every check.
The build runs OpenSCAD with the Manifold backend and then checks the mesh: watertight, one piece, flat on the bed, a real first layer, walls thick enough to print, no overhang past 45 degrees and the size the recipe promised. When a check fails, stone names the one setting to change and prints the command that changes it.
The skill gives each model a card that fits it.
| Tier | Models | What the model does |
|---|---|---|
| S | about 3B and under | picks a recipe and sets its numbers, nothing else |
| M | about 4B to 30B | also picks the look: style, pattern, rounding, text |
| L | frontier models | writes new recipes in OpenSCAD with BOSL2 and the stone library |
Set it with STONE_TIER=S or tier = "S" in stone.toml. stone status prints the card's rules every time, so they come back after a compaction too.
stone new "A keychain of my dog" --image dog.png
stone traces the outline for keychains, cookie cutters and signs, and makes a height map for lithophanes and reliefs. The model never has to look at the picture to use it.
fan_guard, panel, cable_comb, vase, planter, gridfinity_bin, phone_stand, hook, spur_gear, box, keychain, cookie_cutter, sign, lithophane and relief. Every recipe prints without supports and every default and example is built and checked in CI.
You need uv, OpenSCAD 2026 or newer (a nightly build has the Manifold backend) and potrace for pictures.
git clone https://github.com/baney75/smoothstone
cd smoothstone
uv tool install -e .
stone status
stone finds OpenSCAD through STONE_OPENSCAD, then openscad on your PATH, then vendor/openscad/AppRun.
The skill is the skill folder. Link it where your agent looks for skills:
ln -s "$PWD/skill" ~/.claude/skills/smoothstone # Claude Code
ln -s "$PWD/skill" ~/.codex/skills/smoothstone # Codex
ln -s "$PWD/skill" ~/.pi/agent/skills/smoothstone # pi
For a local model on llama.cpp, give it one tool named stone that runs a stone command line, and put skill/tiers/S.md in the system prompt. evals/run.py does exactly this against llama-server and is the harness behind the numbers above.
A recipe is a folder in recipes with recipe.scad and recipe.json. The json lists every setting with its range and unit, the size the part should come out at and examples. uv run python tools/sheet.py <name> builds the defaults and every example through stone and shows the previews. The L card has the full format.
MIT. BOSL2 in vendor/BOSL2 is BSD 2-Clause. The cat, rabbit and tree pictures in examples/images are Material Design Icons under Apache 2.0.