Databricks has released Omnigent, an open-source "meta-harness" that sits above multiple AI coding agents and lets developers swap between Claude Code, OpenAI Codex, Pi, and custom agents without rewriting code. The tool, shipped under the Apache 2.0 license and built in partnership with Neon, addresses a growing friction point: many engineers now juggle four or five different coding agents at once, copying text between tools and losing context across sessions.

What Problem Does Omnigent Actually Solve?

The core insight behind Omnigent is straightforward but powerful. Whether an AI agent is Claude Code, Codex, or a custom model, the user-facing interface is nearly identical: messages and files go in, text streams and tool calls come out. Omnigent standardizes that interface so harnesses become swappable. Instead of re-integrating each tool when you want to switch agents, you change a single line of configuration. This matters because developers often need different agents for different tasks, but switching tools today means abandoning your current session and starting over in a new interface.

The architecture rests on two main components. A runner wraps any agent in a sandboxed session with a uniform API, while a server provides policies and sharing capabilities. One command starts a session in your terminal and simultaneously launches a local web interface at localhost:6767, keeping the same session synchronized across terminal, browser, and mobile.

How to Set Up and Use Omnigent

• Installation: Omnigent requires Python 3.12 or higher, Node.js 22 LTS, and tmux. A single curl command installs everything, then you run "omni setup" to configure model credentials.
• Model Switching: You supply your own models and infrastructure. Omnigent accepts API keys for Claude, ChatGPT, OpenAI-compatible gateways, and Databricks workspaces. The "/model" command switches models mid-session without interrupting your work.
• Custom Agents: Define custom agents using a short YAML file that declares a prompt, harness type (Claude SDK, Codex, Pi, or others), tools, and optional sub-agents. Run any custom agent with a single command.
• Policy Enforcement: Policies use the same YAML approach and stack across three levels: server-wide, per-agent, and per-session. A built-in policy can cap spending with a soft warning at one threshold and a hard limit at another, pausing the agent for approval when spending reaches your configured maximum.

Three Core Capabilities That Set Omnigent Apart

Databricks frames Omnigent around three capabilities designed to solve real developer pain points. Composition means combining models, harnesses, and techniques without rewriting code. You can switch between Claude Code, Codex, Pi, and custom agents with one-line changes instead of rebuilding integrations. Control means stateful, contextual policies that track agent actions and enforce guardrails at the meta-harness layer, not through prompts. One example pauses an agent after every $100 in spending; another requires human approval before the agent runs "git push" after installing a new npm package. Collaboration means sharing live agent sessions by URL so teammates can watch the agent work in real time, chat with it, comment on files, co-drive the session, or fork the conversation.

An OS sandbox called Omnibox underpins these capabilities. It can lock down operating system access and transform network requests. For instance, it can keep your GitHub token hidden from the agent and inject it only in the egress proxy on approved requests, preventing the agent from accidentally exposing secrets.

Practical Patterns Already Emerging

The Databricks team has documented several real-world patterns that Omnigent enables. Polly is a multi-agent coding orchestrator that writes no code itself; instead, it plans work and delegates to coding sub-agents running in parallel git worktrees, with each diff reviewed by a different vendor than the one that wrote it. Debby is a brainstorming partner with two heads, one Claude and one GPT, where every question goes to both models with answers shown side by side, and typing "/debate" makes the heads critique each other before converging. Other patterns include using a frontier advisor model to guide a cheaper open-source worker, having a lead agent orchestrate parallel subagents, or assigning different LLMs to handle planning, search, and code generation in a single workflow.

What's Included and What Isn't

Omnigent ships with several production-ready features. The tool provides one interface to Claude Code, Codex, Pi, and custom agents; sessions reachable from terminal, web, desktop, and phone; policies that track state rather than just allowing or denying actions; live session sharing that replaces copy-pasting between tools; cloud sandboxes on Modal and Daytona that require no local laptop; and an Apache 2.0 open-source license with deployment targets like Fly.io, Railway, and Render.

However, the project is still in alpha and early in its lifecycle. It requires Python, Node.js, and tmux setup, meaning you'll need some command-line familiarity. You bring your own models, infrastructure, and spending budget. Roadmap items like the Omnigent Server MCP (Model Context Protocol) are not yet shipped, and off-network teammates need an always-on deployed server to join sessions.

The timing of Omnigent's release reflects a broader shift in how developers work with AI agents. As multiple coding agents mature simultaneously, the friction of switching between them has become a real productivity bottleneck. By treating agents as interchangeable components, Omnigent removes that friction and lets developers focus on choosing the right tool for each task rather than being locked into a single vendor's ecosystem.