Databricks Open-Sources Omnigent: A Meta-Harness to Unify Your AI Agents

What Is a Meta-Harness?

An agent harness is the scaffolding that turns a model into an agent — giving it the ability to read files, run terminal commands, and call external APIs. Claude Code, Codex, and Pi are all harnesses.

A meta-harness sits one level up: a layer that wraps multiple harnesses simultaneously, so the same policies, sessions, and collaboration tools apply regardless of which underlying agent is running. Databricks describes the meta-harness as the next major abstraction layer in the AI stack, comparable in importance to how cloud infrastructure abstracted individual servers.

Three Core Capabilities

Composition

Omnigent wraps Claude Code, Codex, Pi, OpenAI Agents SDK, Claude Agents SDK, and custom agents behind a single uniform API. Multi-agent workflows — where several agents work on sub-tasks in parallel, or a coordinator delegates to specialists — can be configured without rewriting integration code.

Control

Traditional agent policies are prompt-based, meaning the agent itself enforces them and can be overridden by context drift or adversarial input. Omnigent's policies are stateful and contextual: they track agent actions at the meta-harness layer and can enforce rules like "pause when session spend exceeds $3" or "require human approval before any git push that follows an npm install" — enforced by the harness, not by the model.

An OS sandbox called Omnibox intercepts network requests so agents never see credentials directly. Secrets like GitHub tokens are injected only on approved outbound requests via an egress proxy.

Collaboration

Agent sessions can be shared via URL. Teammates can join a live session, review files, leave comments, co-drive the agent, or fork the conversation — all in real time, without copy-pasting between tools.

Architecture

Omnigent has two components.

Runner: Wraps any agent in a sandboxed session with a uniform API. One command starts a terminal UI and simultaneously launches a local web UI at http://localhost:6767 — the same session, reachable from browser, phone, or desktop app.

Server: Handles policies and sharing, exposing every session over terminal, web app, and API. Deploy to a server with a stable URL (Fly.io, Railway, Modal, Daytona, Render) and sessions become reachable from anywhere, including mobile.

Feature Comparison

Getting Started

Omnigent requires Python 3.12+, Node.js 22 LTS, and tmux. Install in a single command:

uv tool install -q --python 3.12 git+https://github.com/omnigent-ai/omnigent.git
omnigent

Omnigent supports four credential types: first-party API keys (Anthropic, OpenAI), Claude Pro/Max or ChatGPT subscriptions, any OpenAI- or Anthropic-compatible gateway (OpenRouter, LiteLLM, Ollama, vLLM, Azure), and Databricks workspace profiles. Switch models mid-session with the /model command.

Roadmap

Why This Matters

Omnigent is not just a developer convenience. It stakes out the infrastructure layer above every AI coding agent that exists today. The policies, integrations, and session habits that accumulate inside a meta-harness over time create the same kind of durable switching costs that made data platforms sticky — even when the underlying code is open source.

Databricks open-sourced Omnigent two days before its own annual summit keynote. The adoption signal that accumulates this weekend becomes the data Matei Zaharia walks onto the keynote stage with on Monday.