We've built this as part of MUXI, an open-source AI agent infrastructure platform.
The problem: agents need to execute code, not just generate it. Most solutions either give full shell access (terrifying) or sandbox so heavily that nothing useful runs.
skills-rce sits in the middle: a single binary with a standardized OpenAPI-spec'd REST API. You point it at a skill, it runs the code in an isolated subprocess with skill directory caching, and returns the result. Language-agnostic, single endpoint.
It ships as a single binary, but we recommend running it via Docker so execution never touches your host machine. The image is ready to go.
It's intentionally simple. No orchestration opinions, no framework lock-in. Just a code execution service your agents can call.
Apache 2.0. Feedback welcome!