OpenSkyfall

Rebuild CAD assemblies, part by part, with LLMs and skills.

OpenSkyfall is an independent open-source research project inspired by Skyfall's assembly workflow. It is not affiliated with, endorsed by, or an official release of Skyfall AI. It does not include SKY-001 code, weights or API access.

imageIllustrative relay and socket model in the STEP viewer

A native smlibpy demo with 27 placed bodies and 32 saved construction stages. The internal mechanism is illustrative. The drawing-fidelity audit documents incorrect terminal geometry and omitted features. This is not an accurate reconstruction or a manufacturing model.

What we are building

A lightweight agent skill accepts text, images, STL and manual context. It plans a physical-part tree, builds each unique part through smlibpy, then places the parts and exports STEP. The conversation manages requirements and evidence. The side window shows parts, an exploded view, saved construction and the current STEP download. No project-specific model training is required.

What works today

Solid validity, source fidelity, physical decomposition and manufacturing acceptance are separate checks. A valid STEP or an attractive explosion does not establish the others. See architecture and limits.

Run locally

Install uv and Node.js 22 or newer. The main service uses Python 3.14. Native smlibpy runs in a separate Python 3.12 environment.

git clone https://github.com/intgral-ai/open-skyfall.git
cd open-skyfall
npm --prefix web ci
npm --prefix web run build
uv sync --frozen

uv venv --python 3.12 .venv-smlib
uv pip install --python .venv-smlib/bin/python -r requirements-native.txt
export OPEN_ASSEMBLY_SMLIB_PYTHON="$PWD/.venv-smlib/bin/python"
uv run --frozen open-skyfall --workspace jobs serve --port 8766

Open http://127.0.0.1:8766/. On Windows, use the interpreter under .venv-smlib/Scripts/python.exe and set the environment variable accordingly. Availability of a suitable native wheel depends on the platform. The public release check is described in verification.

To try the bounded native example from another terminal:

curl --fail http://127.0.0.1:8766/api/native-cases
curl --fail -X POST http://127.0.0.1:8766/api/native-examples \
  -H 'Content-Type: application/json' -d '{"case_id":"wiring-cover"}'

The response includes a job ID. Open http://127.0.0.1:8766/?job=<id> and wait for the build. This demonstrates an existing example, not reconstruction of a new uploaded object. If smlibpy is unavailable, the service reports that state instead of switching kernels.

The Python module and open-assembly CLI remain as compatibility names.

Use the skill

Read skills/open-skyfall/SKILL.md. A host with local MCP support can launch the server using mcp.json:

uv run --frozen open-skyfall --workspace jobs mcp

The tools are workspace_request, show_assembly and get_reference_image. Keep project context and editing in the chat. The viewer is for model inspection and download. MCP packaging alone does not connect a real ChatGPT account; that requires a compatible host and its deployment configuration.

Manual-driven authoring is demonstrated in the relay/socket case. The original datasheet and user-supplied CAD files are not distributed. Supply your own authorized copy to reproduce a case that requires them.

Contribute

Start with CONTRIBUTING.md. Useful contributions include source-to-CAD checks, native part authoring, documented failures and reproducible examples. The roadmap records the remaining work. No benchmark in this repository establishes a cost or accuracy advantage over SKY-001.

License and references

Project-authored source is MIT. Dependencies, fonts and vendored reference skills keep their own licenses. See LICENSE, NOTICES.md and the pinned upstream references.

The core native dependency is smlibpy. The workflow also draws on text-to-cad and Anatomy. Optional factory intent uses Specify.