OmoiOS

Describe the work. OmoiOS plans the dependency graph, runs a self-supervising swarm of coding agents in isolated sandboxes, and lands your PRs — so you stop being the orchestrator.

  Apache-2.0 · self-hostable · model-agnostic (Claude / Gemini / Codex). Proven end-to-end: one session cloned expressjs/express and ran its full 1,249-test suite inside the spawned sandbox.

  If you find OmoiOS useful, consider giving it a star — it helps others discover the project.

The Problem

AI coding agents are powerful individually, but using them at scale is a mess. You paste a prompt, wait, review, paste another prompt, fix what broke, repeat. There's no dependency awareness, no parallel execution, no structured handoff between tasks. Agents don't know what other agents are doing. When something fails, you're the orchestrator.

OmoiOS fixes this. It reads your existing codebase, generates specs from what's actually there, builds a task DAG with real dependencies, and runs agent swarms across isolated sandboxes until the work is done. A supervisor agent handles merges and keeps everything on track.

This isn't prompt chaining. It's a structured runtime for agent swarms — with dependency graphs, sandboxed execution, active supervision, and code that actually merges.

What Makes This Different

Specs from your actual code
OmoiOS doesn't generate generic plans. It reads your repo — file structure, patterns, dependencies — and generates specs grounded in what exists. The SpecStateMachine runs phases (Explore → Requirements → Design → Tasks) where each phase builds on real codebase context.

DAG-based execution, not a task queue
Tasks form a dependency graph (DependencyGraphService). Nothing executes until its dependencies are met. Critical path analysis determines what runs in…