omnigent-ai/omnigent

Omnigent – a meta‑harness for AI agents

What it is – Omnigent is an open‑source orchestration layer that lets you run, combine, and manage many different code‑generation or chat agents (Claude Code, Codex, Cursor, OpenCode, Hermes, Pi, and any custom agent you describe in YAML). It provides a single CLI (omnigent / omni) and a web‑based UI (also wrapped in a native macOS desktop app) that keep your sessions in sync across terminals, browsers, and phones.

Key capabilities

• Multi‑agent sessions – launch several agents in the same chat, have one review another’s output, or split a task among agents that excel at different subtasks.
• Model‑agnostic – works with API keys, subscription CLIs, or any OpenAI/Anthropic‑compatible gateway (OpenRouter, Ollama, LiteLLM, Azure, etc.) and even Databricks workspaces.
• Cloud sandbox execution – sessions can run in disposable sandboxes on services such as Modal, Daytona, Blaxel, E2B, Kubernetes, NVIDIA OpenShell, Boxlite, and more, so you don’t need a powerful local machine.
• Policy & governance – define policies that pause risky actions, cap spend, or restrict tool usage at the server, agent, or chat level.
• Collaboration – share a live session URL with teammates; they can watch the agent act, take over, or fork the conversation. Multi‑user accounts are supported via a simple auth toggle.
• Cross‑device continuity – start a chat in a terminal, continue it in the browser, and pick it up on a phone; all messages, files, and sub‑agents stay synchronized.
• Extensible harnesses – the system ships harnesses for the major agents and lets you add your own via the SDK or by describing an ACP‑compatible command in YAML.

Typical workflow

1. Install – a one‑line script installs the Python package, required tools (uv, git, Node, tmux, etc.) and optional extras (model or sandbox providers).
2. Start a session – omnigent launches a local server, opens the UI at http://localhost:6767, and starts a default agent (or a specific one like omnigent claude).
3. Pick a model – omnigent setup lets you add API keys, subscriptions, or gateway URLs and set defaults per agent.
4. Run agents – use built‑in examples (polly, debby, deep‑research) or run your own YAML‑defined agents with omnigent run <path>. You can also switch models mid‑chat via the /model command.
5. Deploy – a single docker compose up (or one‑click deploy on Render, Railway, Fly.io, Cloudflare, etc.) runs the server publicly, enabling phone access and managed cloud sandboxes.
6. Collaborate – enable multi‑user mode, invite teammates, and work together on the same session.

Who it’s for – developers, data scientists, or teams that want a unified interface for experimenting with many LLM‑based coding or reasoning agents, enforce safety policies, and share work without rewriting integration code for each provider.

Status – early‑alpha (v0.x) but actively maintained, with a public Discord for community support.