threecrate

threecrate

Fast 3D point cloud and mesh processing in Rust and Python, with an agent skill that lets Claude Code, Codex, Gemini CLI and others work on your scans.

Ask your agent to "make my scan printable" and it cleans the scan, closes the holes, adds a base, checks the result with a picture and saves an STL. No Blender, no MeshLab, no guessing at parameters.

A real, unedited Claude Code run with the threecrate skill, replayed. More in the demo.

Installation (30-second setup)

1. Get the skill

claude plugin marketplace add rajgandhi1/threecrate
claude plugin install threecrate@threecrate

Or type /plugin marketplace add rajgandhi1/threecrate, then /plugin install threecrate@threecrate, as two prompts.

codex plugin marketplace add rajgandhi1/threecrate
codex plugin add threecrate@threecrate

Start a new thread after installing.

gemini extensions install https://github.com/rajgandhi1/threecrate

Copy the skills/threecrate folder into your agent's skills folder. INSTALL.md has the exact commands for each agent.

claude mcp add threecrate -- uvx threecrate-mcp

threecrate-mcp gives the agent tools such as inspect_mesh, repair_for_print and render, which hands it a picture of the result.

The skill needs python 3.8 or newer on your PATH. The agent installs threecrate into a virtual environment the first time.

2. Ask in plain words

• "Make table_scan.ply printable."
• "Is part.stl ready to print?"
• "Clean up this scan and drop the floor."
• "Align these two scans."
• "How tall is the pole in street.pcd?"

3. That's it

Want to try it first? samples/ has small scans with known answers, and the demo shows a full run.

Why threecrate

#1: Scans don't print

The problem. A phone or LiDAR scan comes out with stray points, the table it stood on, holes, no base and flipped faces. The slicer rejects it, and fixing it means learning MeshLab or Blender.

The fix. repair_for_print takes a point cloud or a mesh and returns a watertight STL, with a report of every step. analyze_mesh tells you what is wrong before you start.

#2: Agents can't see 3D

The problem. An agent can run a mesh tool, but it can't tell whether the result looks right. It reports success on a blob.

The fix. render draws any cloud or mesh from five views into one PNG, with no GPU. Holes and inside-out faces show orange. The skill makes the agent look at the before and after picture before it reports.

#3: Parameters are guesswork

The problem. Every voxel size, radius and threshold depends on the scan's units and density. Wrong values delete real points or do nothing.

The fix. The skill inspects each scan first and sizes every parameter from its point spacing. One guide per task keeps the agent on a known path: clean, align, make printable, measure.

#4: 3D libraries are slow or heavy

The problem. Open3D and PCL are large C++ builds. Pure Python is too slow for millions of points.

The fix. threecrate is Rust with a pip install. It is 1.5x to 3.4x faster than Open3D on reading, downsampling, normals and ICP across TUM, KITTI and nuScenes frames, with the same ICP accuracy (measured on Windows; in a Linux container the two are close). GPU code runs on any card through wgpu. Full numbers in docs/benchmarks.md.

Reference

Python

pip install threecrate

import threecrate as tc

scan = tc.read_point_cloud("scan.ply")  # or tc.read_mesh("part.stl")
mesh, report = tc.repair_for_print(scan, remove_floor=True)
print(report["steps"], report["after"]["printable"])
tc.render([scan, mesh], "before_after.png")  # orange shows holes and inside-out faces
tc.write_mesh(mesh, "part.printable.stl")

All functions: Python API reference.

Rust

[dependencies]
threecrate = "0.12.1"

use threecrate::prelude::*;

let cloud = read_point_cloud("scan.ply")?;
let cloud = voxel_grid_filter(&cloud, 0.05)?;
let normals = estimate_normals(&cloud, 10)?;
let mesh = auto_reconstruct(&normals)?;
write_mesh("output.obj", &mesh)?;

Crate	What it does
threecrate-core	Point, PointCloud, TriangleMesh, Transform3D, mesh health checks
threecrate-algorithms	Filtering, ICP, NDT, global registration, segmentation, normals, FPFH/SHOT, mesh boolean, smoothing
threecrate-gpu	GPU filtering, segmentation, ICP, normals, nearest-neighbor, TSDF, real-time rendering (wgpu)
threecrate-io	PLY, OBJ, STL, PCD, XYZ/CSV, LAS/LAZ*, E57*, streaming and memory-mapped; PNG previews without a GPU
threecrate-reconstruction	Poisson, ball pivoting, alpha shapes, Delaunay, marching cubes, MLS, auto-select, repair for 3D printing
threecrate-simplification	Quadric error, edge collapse, clustering, progressive mesh
threecrate-visualization	Interactive viewer: orbit, pan, zoom, GPU-accelerated

* opt-in feature flags. See installation and feature flags.

Feature	threecrate	Open3D	PCL
Language	Rust + Python	Python (C++ core)	C++
pip install	✅	✅	❌
Memory safety	✅ Rust	❌	❌
GPU compute	✅ wgpu, any GPU	✅ CUDA	Partial
Global registration	✅ FPFH + RANSAC	✅	✅
Surface reconstruction	✅ 6 algorithms	✅	✅
Repair for 3D printing	✅ one call	Partial	❌
Agent skill and MCP server	✅	❌	❌
Streaming I/O	✅ PLY/OBJ/XYZ	❌	❌
E57 support	✅ opt-in	❌	❌

Docs

• Crate reference
• Examples
• Benchmarks
• API docs on docs.rs

Contributing

Contributions are welcome: algorithms, Python bindings, new formats, docs.

• ROADMAP.md: where we're ahead, where we trail, and what's next
• CONTRIBUTING.md: setup and guidelines
• Open issues: look for good first issue
• GitHub Discussions: questions and ideas

License

MIT OR Apache-2.0.