Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/omnigent-ai/omnigent/llms.txt

Use this file to discover all available pages before exploring further.

Omnigent is an open-source meta-harness and declarative runtime framework for AI agents. Instead of being locked into one agent runtime, Omnigent gives you a common layer over Claude Code, Codex, Pi, and the agents you write yourself — so you can swap or combine harnesses without rewriting anything, govern every action with policies, and collaborate with teammates in real time on the same live session from any device, including your phone.
Omnigent is currently in alpha. The API and YAML schema may change between releases. Pin your install to a specific version in production workloads and review the changelog before upgrading.

Unified Harnesses

Run Claude Code, Codex, Pi, and custom YAML-defined agents side by side in the same session. Switch harnesses with a single flag — no rewrites required.

Policies & Governance

Define server-wide, per-agent, or per-session rules that approve, block, or pause actions — shell commands, file edits, token spend — before they execute.

Real-Time Collaboration

Share a live session link, co-drive from a teammate’s terminal, or fork a conversation to continue independently from any point in the history.

Cloud Sandboxes

Run sessions in disposable Modal or Daytona sandboxes launched from the CLI, or let the server provision them automatically as managed hosts.

​
How Omnigent Works

Omnigent is built around three cooperating layers. The server is a FastAPI/Uvicorn process that owns session state, the policy engine, the user database, and the web UI. It runs locally (omnigent server start) or on any always-on host you deploy. The runner is the process that actually executes an agent session — it loads your YAML, resolves credentials, wires up tools, and streams events back to the server in real time. The harness is the thin adapter the runner delegates model calls to: claude-sdk, codex, openai-agents, pi, or a custom harness you provide. Because the harness is just a named plugin, swapping from Claude Code to Codex (or running both in parallel as sub-agents) is a one-line change in your agent YAML.

​
Who It’s For

Omnigent is designed for four kinds of people working with AI agents:
  • Developers who want to run Claude Code, Codex, or Pi from a single CLI, switch models without reconfiguring their environment, and keep sessions accessible from any device.
  • Agent authors who want to define multi-agent workflows in YAML — with prompts, tools, sub-agents, and reviewers — and iterate quickly without writing orchestration boilerplate.
  • Platform and security teams who need to govern what agents can do: cap token spend, require approval before shell commands run, and audit every tool call across a shared server.
  • Teams who want to co-develop with AI: sharing live sessions, co-driving on the same running agent, or forking a conversation so two people can explore different directions simultaneously.

​
Key Capabilities

  • Multi-harness sessions — Claude Code, Codex, Pi, and openai-agents run as first-class harnesses; combine them in a single YAML as a supervisor and sub-agents.
  • Declarative agent YAML — define your agent’s prompt, harness, tools (Python functions or nested sub-agents), and policies in one file; run it with omnigent run.
  • Policy engine — stack server-wide, per-agent, and per-session policies using built-in handlers for shell approval, spend caps, and tool-call limits; write custom handlers in Python.
  • Real-time web UI — every session gets a live browser view at http://localhost:6767, built for desktop and mobile, with sub-agent trees, terminal panels, and file diffs.
  • Cloud sandboxes — launch disposable Modal or Daytona sandboxes from the CLI or let the server provision them per session; no laptop needs to stay online.
  • Multi-user collaboration — invite teammates via single-use links, share live sessions, co-drive with omnigent attach, or fork with omnigent run --fork; SSO via OIDC optional.
  • Any model or gateway — first-party API keys, Claude Pro/ChatGPT subscriptions, or any OpenAI/Anthropic-compatible gateway (OpenRouter, LiteLLM, Ollama, vLLM, Azure).
  • macOS desktop app — wraps the same web UI in a native window with OS notifications and a dock badge; download at omnigent.ai.

​
Next Steps

Quickstart

Install Omnigent and run your first agent session in under five minutes.

Architecture

Dive deeper into the server, runner, and harness model.

Write Your Own Agent

Author a custom agent in YAML with tools, sub-agents, and a harness of your choice.

Policies

Govern every tool call with built-in and custom policy handlers.

Build docs developers (and LLMs) love

Get started for freeTalk to us