Hugging Face Spaces

Orbit is a self-hosted, MIT-licensed harness that turns AI coding-agent runs into closed, verifiable loops. The core workflow: Orbit selects a task from a dependency-ordered backlog, hands it to a configured agent, runs the agent’s output through tests, lint, and type checks, then writes four structured artifacts — a result JSON, an evaluation rubric, a review recommendation, and a human-readable progress log. If the agent cannot pass validation, the orbit does not close. No task is marked complete on assertion alone.

The standout design decision is agent-neutrality. Orbit does not care which coding agent sits behind the contract — only that it speaks JSON over CLI. That means you can run the same task against Claude, Codex, and Cursor, collect the evaluation and diff artifacts from each run, and compare structured output instead of informal impressions. For teams evaluating which agent to standardize on, that is the difference between a reproducible experiment and a gut call.

Orbit fits cleanly in two scenarios: self-healing repositories where failing tests or lint issues are fed as tasks and proof-of-work is required before merge, and CI/CD pipelines where agent-driven changes need an auditable record a human reviewer can inspect. It does not fit when you need a team-facing dashboard, multi-agent parallelism managed by the harness itself, or adapters pre-built for anything beyond the demo’s replay path — the docs describe Orbit as intentionally small, and the contribution guide makes clear that adapters and mission templates are community contributions, not bundled defaults.

The deterministic replay demo runs without an API key using a MOCK flag, which means the validation loop and artifact schema are inspectable before you connect a live agent. Setup is a git clone, a virtual environment, and a shell script — no external service dependencies for the replay path.