Databricks Open-Sources Omnigent: A Meta-Harness That Connects, Controls, and Shares AI Agents on Claude Code, Codex, and Pi

Information bricks removed He is omnipotentopen source ‘meta-harness’ for AI agents. The project runs under the Apache 2.0 license. The Databricks AI team builds on Neon.

A harness is a wrap model that turns into an agent. Claude Code, Codex, and Pi harnesses. Omnigent sits one level above them. Treats each harness as a flexible part of a larger system.

Many developers are now running four or five agents at once. They copy text between coding agents, search tools, Docs, and Slack. Each harness understands only its own times. Omnigent adds a shared layer to live design, control, and collaboration.

What is Omnigent

Omnigent is a common interface on top of command line agents and SDK agents. It wraps terminal coding agents such as Claude Code, Codex, and Pi. It also wraps SDKs like OpenAI Agents and Claude Agents SDK.

The design relies on one observation. However the harness calls its model internally, the user interface is the same. Messages and files are incoming. Text streams and tool calls are coming out. Omnigent sets up that interface so the harnesses can be swapped.

He provides models and infrastructure. Omnigent runs agents up. It can coordinate several of them as alternating workers under one orchestra.

How Omnigent works

The architecture has two parts. The runner wraps any agent in a sandboxed session with the same API. The server provides policies and sharing. The server exposes every session over the terminal, application, and web APIs.

A single command starts a session in your terminal. It also introduces a local web UI to the localhost:6767. The same time appears in the browser or on the phone. Messages, sub-agents, terminals, and files are always synchronized.

CLI includes under two names, omnigent again omni. They are flexible. On first launch, it finds model credentials that are already on your site.

Design, Control, and Cooperation

The Databricks team is completely split into three skills:

• Formation it means combining models, harnesses, and techniques without rewriting the code. Switch between Claude Code, Codex, Pi, and custom agents with one-line changes.
• Take control means contextual policies. They track agent actions and enforce monitoring rules at the meta-harness layer, not through commands. One example stops an agent after every $100 they spend. Others require a person’s consent to git push when the agent installs a new npm package.
• Cooperation means sharing live agent sessions via URL. Teammates watch the agent in action and chat with it in real time. They can comment on files, run a joint session, or reply to a chat.

The OS sandbox, called Omnibox, supports this. It can block OS access and modify network requests. For example, it can keep your GitHub token hidden from the agent. The token is only injected into the egress proxy for authorized requests.

Use Cases and Examples

Two examples of agents go with the repository:

• Polly is a multi-agent coding orchestrator. It doesn’t write the code itself. It organizes, and agents work on writing sub-agents in parallel git worktrees. Each is a different approach to the reviewer from a different vendor than the author. You combine the result.
• Debbie he is a two-headed negotiating partner. One head is Claude, the other is GPT. All questions go to both, and the answers are shown on the side. Kind of /debate and heads criticize each other before meeting.

Other active patterns follow a similar pattern. The boundary consultant model can guide a cheap open source worker. A lead agent can organize corresponding subagents. Different LLMs can handle editing, searching, and coding in a single flow.

Interactive Concept Demo

The Marktechpost team has created an interactive demo (below) that lets you experience Omnigent’s meta-harness workflow for yourself. You choose the job of Polly orchestrator, which organizes and sends it to three sub-agents: Claude Code, Codex, and Pi which run in parallel and broadcast their steps live. The session cost meter increases as it runs, and two policy changes reflect the performance of Omnigent’s control layer: the cost budget limits the activity to $3.00 for your consent, and the context policy limits a git push that follows npm installation until you enable it. Once the sub-agents are done, each diff is reviewed by a different vendor than the one that wrote it, and is marked as ready to merge. Terminal, web, and mobile tabs show the same session that stays in sync across all links. Figurative modeling, no live models are called.

Omnigent vs Single Harness

Getting started

Omnigent requires Python 3.12+, Node.js 22 LTS, and tmux. One command installs everything:

curl -fsSL install.sh | sh

Then set the model information:

Omnigent accepts four types of authentication. They are a first person API key and a subscription to Claude or ChatGPT. Others are OpenAI- or Anthropic-compatible gateways and Databricks workspaces. I /model command changes models during a session.

A custom agent is a short YAML file. Announces information, harness, tools, and small agents can choose for you.

name: my_agent
prompt: You are a helpful data analyst.

executor:
  harness: claude-sdk          # or: codex, codex-native, claude-native, openai-agents, pi

tools:
  researcher:
    type: agent
    prompt: Search for relevant information and summarize it.

Run it with one command:

omnigent run path/to/my_agent.yaml

Policies use the same YAML method. These builtin caps spend with a gentle warning first:

Policies use the same YAML approach. This builtin caps spend with a soft warning first:

policies:
  budget:
    type: function
    handler: omnigent.policies.builtins.cost.cost_budget
    factory_params:
      max_cost_usd: 5.00          # hard spend cap
      ask_thresholds_usd: [3.00]  # soft warning on the way

Policies stack at three levels. They are server-wide, per-agent, and per-session. Strict time rules are checked first.

Powers and Limitations

Power

• A single interface to Claude Code, Codex, Pi, and custom agents
• Sessions are accessible on terminal, web, desktop, and phone
• Policies that track status, not just allow or deny
• Live session sharing replaces copying between tools
• Cloud sandboxes in Modal and Daytona do not require a local laptop
• Apache 2.0 license with terms like Fly.io, Railway, and Render

Limitations

• The project is alpha and at the beginning of its life cycle
• Requires Python, Node.js, and tmux setup
• You bring your own models, infrastructure, and spending
• Roadmap items such as the Omnigent Server MCP have not been shipped yet
• Off-network partners need a dedicated server to join

Check out Repo, start immediately, again Technical details. Also, feel free to follow us Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to Our newspaper. Wait! are you on telegram? now you can join us on telegram too.

Need to work with us on developing your GitHub Repo OR Hug Face Page OR Product Release OR Webinar etc.? contact us