Databricks has released Omnigent, an open-source meta-harness for AI agents that sits above individual coding agents and SDKs, providing a unified layer for composition, control, and collaboration. The project, built by the Databricks AI team in collaboration with Neon, ships under the Apache 2.0 license and aims to solve a problem many engineers now face daily: juggling four or five different agent harnesses—Claude Code, Codex, Pi, and others—while manually copying text between terminals, search tools, documentation, and Slack. Each harness understands only its own sessions. Omnigent adds a shared layer where those sessions can be composed, governed, and shared without leaving the developer’s flow.
What Is the Omnigent Meta-Harness
A harness is the software wrapper that turns a language model into an agent. Claude Code, Codex, and Pi are all harnesses. Omnigent sits one level above them. It treats each harness as an interchangeable component within a larger system. The core observation driving the design is that however a harness calls its model internally, the user-facing interface is fundamentally the same: messages and files go in, text streams and tool calls come out. Omnigent standardizes that interface so that harnesses become swappable without re-integration work. Users supply their own models and infrastructure; Omnigent runs the agents on top, coordinating several of them as interchangeable workers under a single orchestrator.
How Omnigent Works: Architecture and Interface
The architecture comprises two main parts. A runner wraps any agent in a sandboxed session that exposes a uniform API. A server provides policies and session sharing. The server makes every session available over the terminal, a local web application, and web APIs. A single command starts a session in the terminal and simultaneously launches a local web UI at localhost:6767. That same session remains accessible from a browser or a phone, with messages, sub-agents, terminals, and files kept in sync across every interface. The CLI installs under two interchangeable names—omnigentcodecodecodecodecode and omnicodecodecodecodecode—and on first run it detects model credentials already present in the user’s environment.
Composition, Control, and Collaboration
The Databricks team frames Omnigent around three core capabilities. Composition means combining models, harnesses, and techniques without rewriting code. Switching between Claude Code, Codex, Pi, and custom agents requires a one-line change. Control refers to stateful, contextual policies that track agent actions and enforce guardrails at the meta-harness layer rather than through fragile prompt instructions. One example policy pauses an agent after every $100 it spends. Another requires human approval before a git push can proceed when the agent has just installed a new npm package. Collaboration means sharing live agent sessions by URL. Teammates can watch an agent work, chat with it in real time, comment on files, co-drive the session, or fork the conversation. An operating-system sandbox called Omnibox underpins this layer. It can lock down OS access and transform network requests—for instance, keeping a GitHub token hidden from the agent and injecting it only in the egress proxy on approved requests.
Use Cases: Polly and Debby
Two example agents ship with the repository. Polly is a multi-agent coding orchestrator that writes no code itself. It plans a task, then delegates work to coding sub-agents running in parallel git worktrees. Each diff is routed to a reviewer from a different vendor than the writer, and the result is marked ready to merge. Debby is a brainstorming partner with two heads: one Claude, one GPT. Every question goes to both models, with answers shown side by side. Typing /debatecodecodecodecodecode prompts the two heads to critique each other before converging. Other practical patterns follow the same shape: a frontier advisor model guiding a cheaper open-source worker, a lead agent orchestrating parallel sub-agents, or different LLMs handling planning, search, and code generation within a single flow.
Omnigent vs. a Single Harness
The difference between working with a single harness and working with the Omnigent meta-harness is substantial. A single harness like Claude Code supports one agent and allows model swapping only inside that harness; switching costs involve re-integration per tool, and governance relies on allow-or-deny lists that are often prompt-based. Omnigent makes Claude Code, Codex, Pi, SDKs, and custom agents interchangeable with a one-line change. Its sessions are accessible from terminal, web, desktop, mobile, and APIs simultaneously. Policies are stateful and contextual rather than simple allow-deny rules. Cost control uses budget policies that pause at set thresholds. Collaboration happens through live shared sessions that can be co-driven and forked, rather than through copy-pasting between tools. The Omnibox sandbox adds an OS-level sandbox plus egress-proxy secret injection, and the project supports disposable cloud sandboxes on Modal and Daytona.
Getting Started with Omnigent
Omnigent requires Python 3.12 or later, Node.js 22 LTS, and tmux. A single command installs everything:
curl -fsSL https://omnigent.ai/install.sh | shpreprepreprepre
The system accepts four credential types: a first-party API key and a Claude or ChatGPT subscription, an OpenAI- or Anthropic-compatible gateway, and a Databricks workspace. The /modelcodecodecodecodecode command switches models mid-session. A custom agent is defined in a short YAML file that declares a prompt, a harness, tools, and optional sub-agents. Running it requires one command: omnigent run path/to/my_agent.yamlcodecodecodecodecode. Policies use the same YAML approach and can be stacked across three levels—server-wide, per-agent, and per-session—with stricter session rules checked first.
Strengths and Limitations
The project’s strengths include a single interface to multiple harnesses, sessions reachable from terminal, web, desktop, and phone, stateful policies rather than simple allow-deny rules, live session sharing that replaces copy-pasting, cloud sandbox support on Modal and Daytona, and an Apache 2.0 license with deployment targets including Fly.io, Railway, and Render. Limitations are equally important to note. The project is alpha and early in its lifecycle. It requires Python, Node.js, and tmux setup. Users bring their own models, infrastructure, and spend. Roadmap items such as the Omnigent Server MCP are not yet shipped. Off-network teammates need an always-on deployed server to join sessions.
Who Should Try This Now
For developers and teams already working with multiple AI coding agents and feeling the friction of managing them separately, Omnigent offers a practical, open-source way to unify those workflows under a single governed session. The project is ready to test today: install it, define a custom agent in a YAML file, and run a session that spans Claude Code, Codex, and Pi simultaneously from one terminal. The interactive concept demo available on the project site provides a hands-on feel for the orchestrator workflow, budget policies, and cross-vendor review process. As the project matures toward production readiness, its approach to stateful policy enforcement and live session sharing points toward a broader industry shift from single-agent tools to multi-agent orchestration platforms.
