Omnigent
A meta-harness for building and running AI agents
About Omnigent
Omnigent is a meta-harness: a common layer that sits above Claude Code, Codex, Pi and agents you write yourself, so you can swap or combine them without rewriting the code around them. Most teams running agents seriously end up with the same problem, which is that spend limits, model routing, sandboxing and credential handling get reimplemented once per harness, in slightly different and slightly wrong ways. Omnigent moves those concerns up a level and makes them policy rather than per-tool configuration.
The architecture is two pieces. A runner wraps any agent in a sandboxed, uniform session. A server sits on top adding contextual policies and shared history, and exposes every session over the terminal, the web, a native desktop app, mobile and a REST API. The policies are stateful rather than static: spend caps that track cumulative usage, model routing, and risk-based escalation that can push a decision to a human. The sandbox restricts filesystem and network access and hides credentials from the agent while brokering access to them, which is what makes running an agent in unattended mode a defensible choice rather than a gamble. It ships two built-in agents to demonstrate the model - Polly, a coding orchestrator, and Debby, a model debate - and custom agents are defined in YAML. It is built by the Databricks AI team and Neon, which is a meaningful signal for a project this young, but the project's own homepage says plainly that Omnigent is alpha and built in the open. At v0.11.0 the API will move, and you would be inserting an alpha layer between yourself and tools that already work.
Install it with the shell script from the project site, or through uv, pip or Homebrew, and start a session. Whatever harness you point it at - Claude Code, Codex, Pi or your own agent - the runner wraps it in the same sandboxed session shape, which is what makes switching between them a one-line change instead of a rewrite.
Policies are attached rather than coded in, so a spend cap, a routing rule or an escalation threshold applies across whichever agent is running underneath. Because the server holds the session and its history, the same live session is reachable from the terminal, a browser, the macOS app, iOS, Android or the REST API. Sessions can be shared by URL with their full history, so a colleague can watch a run in progress and steer it, which is a different collaboration model from pasting transcripts after the fact.
- •Harness Composition - Combine or switch between Claude Code, Codex, Pi and custom agents with one-line changes, so the choice of harness stops being a lock-in decision
- •Contextual Policies - Stateful spend caps, model routing and risk-based escalation applied as policy across every agent rather than configured per tool
- •Secure OS Sandbox - Restricts filesystem and network access and hides credentials from the agent while brokering access to them, which is what makes unattended runs defensible
- •Uniform Session Model - A runner wraps any agent in the same sandboxed session shape, and a server adds policy enforcement and shared history on top
- •Every Surface - The same live session over terminal, web, native macOS app, iOS, Android and a REST API
- •Real-Time Collaboration - Share a live agent session by URL with its full history, so a team can review and steer a run while it happens
- •Built-In and Custom Agents - Polly (a coding orchestrator) and Debby (a model debate) ship as examples, and your own agents are defined in YAML
Teams running more than one agent harness who are tired of maintaining the same guardrails twice, and anyone who needs spend and risk controls to be enforced centrally rather than trusted to each tool's own settings. The collaboration surface suits pairs or teams who want to watch and steer a long agent run together instead of trading transcripts, and the mobile and REST access matter if you supervise runs away from your desk.
It is not for you if you run one harness and are happy with it, because the abstraction only pays for itself across several. It is also not for anyone who needs stability: the project calls itself alpha, and a wrapper's maintenance burden is the sum of everything it wraps, so it depends on Claude Code, Codex and Pi continuing to fit its model. There is no commercial tier and no support commitment behind it - feedback goes to Discord - so treat it as something to evaluate and contribute to rather than to put under production workloads today.
