Omnigent: An Open-Source Meta-Harness for Orchestrating Claude Code, Codex and Other Agents

Omnigent is an open-source meta-harness for AI coding agents. It is not another coding agent.

It sits above Claude Code, Codex, Cursor, OpenCode, Hermes, Pi and agents you write yourself, and it adds one common layer for sessions, policies, sandboxes and collaboration. The project uses the Apache 2.0 license, ships on PyPI, and the README marks it as alpha.

The problem it targets

Coding agents have multiplied. Each one has its own terminal UI, permission model, session format and billing path. A team that uses several of them must switch between experiences, and it is hard to apply one set of rules to all of them.

Omnigent separates the agent itself from the environment around it. The agent reasons and edits code. Omnigent handles the session, the permissions, the isolated runtime and the access from many devices. When you swap or combine agents, you do not rebuild the surrounding tooling.

Core architecture

Based on the project's README, Omnigent has four main parts. First, a shared session layer. Messages, sub-agents, terminals and files stay in sync across clients. You can start a task in the terminal, continue in the browser, and check progress on your phone. A native macOS desktop app is also offered.

Second, multi-agent orchestration. One session can mix several coding agents. One agent can review the work of another, or a task can be split across agents that each have different strengths. Custom agents are defined in YAML. Third, a model access layer. You can use a first-party API key, a Claude or ChatGPT subscription, or any compatible gateway. The README calls all of these first-class. Fourth, a policy engine and a sandbox runtime, covered below.

Policy and governance

Policies are the feature most relevant to enterprise use. The README lists three typical uses: pause for human approval before a risky action, cap spend, and limit which tools an agent can reach. A policy can apply to the whole server, to one agent, or to a single chat. The scopes layer from broad to narrow. A team could block a class of tools server-wide, then relax the rule for one trusted agent. Tool approvals for external agents such as Devin appear as chat approval cards, which shows the approval flow lives inside the common interface.

One caveat. The README excerpt does not describe how the policy engine is built. It also does not say whether it can intercept every tool call made inside a third-party agent's own CLI. Check the source and the official docs before you rely on it as a hard security boundary.

Cloud sandboxes and managed hosts

Omnigent can run sessions in disposable cloud sandboxes. The listed providers are Modal, Daytona, Blaxel, Islo, E2B, Gensee, CoreWeave, Kubernetes, OpenShell, Boxlite, microsandbox and Databricks. You can launch them from the CLI, or the server can provision one per session.

The README calls this a managed host. The practical gain is clear: no laptop needs to stay awake, and risky or long tasks run in an environment you can throw away. Each sandbox provider is an optional extra, such as modal, daytona, e2b or kubernetes, so the core install stays small.

Collaboration

A session can be shared.

Teammates can chat with your agent and watch it work live, co-drive it on your machine, or fork the conversation and continue on their own. This turns an agent from a personal tool into a shared resource that others can review.

Install and setup

The recommended path is one command: download scripts/install_oss.sh with curl and run it. You can also use uv tool install omnigent, pip, or Homebrew (omnigent-ai/tap/omnigent). Omnigent needs Python 3.12 or newer.

The coding-agent CLIs need Node.js 22 LTS with npm, and the web UI needs pnpm. Extras cover model providers (databricks, bedrock, vertex), SDK harnesses (antigravity, copilot, cursor, agents-sdk), and storage and memory (s3, hindsight). The optional Devin CLI has its own installer.

Performance and cost

The material we reviewed contains no benchmarks, and no latency or cost comparisons. We therefore give no performance numbers. The cost structure is clear, though.

Omnigent is free open-source software. The real spend comes from model calls and from cloud sandbox billing. The spend-cap policy is the tool for controlling that. An extra orchestration layer adds processes and network hops, and its overhead needs independent measurement from the community.

Ecosystem impact

For developers, Omnigent lowers the cost of moving between coding agents. Using one agent to write and another to review becomes a configuration task, not an engineering project.

For enterprises, shared approvals, spend control and sandbox isolation form the governance layer that is often missing when agents enter production workflows. Support for newer open agents such as Hermes and Pi shows that the project tracks the open agent ecosystem.

Limits and challenges

The project is alpha. Interfaces and behavior may change, so do not place it in a critical production path yet. A meta-harness must follow the release cycle of every agent below it, so adapter upkeep never ends.

A common abstraction can hide features that are unique to one agent. The strength of a policy depends on how deeply it integrates with each agent, so test it yourself. Finally, the toolchain is wide: Python, Node and pnpm all appear in the prerequisites.

What to watch next

Watch for more agent and sandbox integrations, published benchmarks and cost data, a finer policy language, and maturing desktop and mobile apps. If Omnigent leaves alpha and shows stable gains from cross-agent orchestration, it could become a general control plane for multi-agent development.