Databricks has open sourced Omnigent, a meta harness designed to compose, govern, and share AI agents across popular models like Claude Code, Codex, and Pi. The project is released under the Apache 2.0 license and was developed by the Databricks AI team using Neon.
Omnigent serves as an intermediary layer between AI harnesses and their underlying models. It standardizes the interface so that various agents can be swapped in and out seamlessly, allowing users to manage multiple agents more efficiently. Engineers now commonly use several agents simultaneously, but Omnigent introduces a shared environment where composition, control, and collaboration take place.
This meta-harness wraps command-line coding agents such as Claude Code, Codex, and Pi, as well as SDKs like OpenAI Agents and the Claude Agents SDK. The design leverages the observation that despite differing internal implementations, all harnesses present the same user-facing interface: messages and files go in; text streams and tool calls come out.
Omnigent’s architecture consists of two main components:
1. Runner: This wraps any agent in a sandboxed session with a uniform API.
2. Server: Provides policies and sharing features, exposing every session via the terminal, app, and web APIs.
To start using Omnigent, one command initiates a session in your terminal, launching a local web UI at localhost:6767. The same session appears across various interfaces browser, phone, or desktop. Messages, sub agents, terminals, and files remain synchronized.
Databricks frames Omnigent around three core capabilities:
1. Omnibox: An OS sandbox that can lock down access to certain resources.
2. Policies and Sharing: These control aspects of the session, such as cost management and contextual restrictions.
3. Interchangeability: Harnesses become swappable components within a single orchestration layer.
A demo is available on the MarkTechPost website, illustrating how Omnigent manages multiple agents in parallel. The demo uses Polly as an orchestrator, delegating tasks to Claude Code, Codex, and Pithat running in parallel, with live streaming of steps. Policies like cost budgeting and contextual gating are also demonstrated.
To set up Omnigent, you need Python 3.12+, Node.js 22 LTS, and tmux. Installation is straightforward:
“`bash
git clone https://github.com/databricks/omnigent.git
cd omnigent
pip install -r requirements.txt
npm install
./install.sh
“`
For model credentials, Omnigent supports four types: a first-party API key, a Claude or ChatGPT subscription, an OpenAI or Anthropic compatible gateway, and a Databricks workspace. Models can be switched mid session using the `model` command.
Custom agents are defined in short YAML files that declare prompts, harnesses, tools, and optional sub-agents. Policies also follow this structure to enforce rules at different levels (server-wide, per-agent, or per-session).
Thinking about building an AI product?
Get in Touch