Open source · Alpha

Omnigent, a meta-harness for AI agents

A common layer over Claude Code, Codex, Pi, and the agents you write yourself: swap or combine harnesses without rewriting, keep them in check with policies and sandboxing, and pick the same live session back up anywhere, on your phone or with a teammate.

$ pip install omnigent

Desktop app: Download for macOS

Features

  • Polly & Debby: built-in multi-AI agents (a coding orchestrator and a model debate). Build your own in YAML.
  • Contextual Policies: stateful spend caps, model routing, and risk-based escalation.
  • Flexible OS sandbox: restrict filesystem, network, and env to run agents in YOLO mode safely.

What Omnigent gives you

Omnigent mascots stacked in layers to illustrate composability.

Composition

The best results increasingly come from combining models and techniques: an advisor model here, a router there, two harnesses debating to raise quality. And the “best” harness keeps changing as models evolve. How do you compose those pieces, or switch harnesses, without rewriting everything?

Omnigent's mascot supervising smaller agents playing in a sandbox, illustrating the OS-level sandbox that isolates every agent.

Control

Agents are held back today because we don’t fully trust them. Can you let an agent run at full speed and still guarantee it won’t delete prod, leak sensitive data, or quietly burn $1,000 on a bad idea? Right now the answer is usually “watch it closely,” which defeats the purpose.

Two Omnigent mascots high-fiving to celebrate collaboration.

Collaboration

You just spent two hours getting something right with an agent, whether that’s a refactor or a market analysis. Why can’t you invite a colleague into that same session to review it, or hand it off, the way you’d share a doc?


How it fits together

A runner wraps any agent in a sandboxed, uniform session. A server adds policies and a shared history, and exposes every session over the terminal, the web, and a REST API.

Omnigent architecture: CLI agents and custom agents run through a runner that sandboxes them, then a server that adds policies and history, reachable from a terminal, the web, and a REST API.

Try it

Install, then start an agent in your terminal. A local web UI opens too, for the browser or your phone.

pip install omnigent
omni                     # Polly, the default multi-agent orchestrator
omni claude              # or launch a specific harness
omni run my-agent/       # or run your own agent

Read the installation guide, then write your own agent in a few lines of YAML.


Early days

Omnigent is alpha and built in the open. Try it, break it, and tell us on Discord.