Omnigent Framework Addresses AI Agent Governance Gap
The new open-source Omnigent framework acts as a meta-harness for multiple AI coding agents, aiming to standardize credential handling and command execution to improve oversight and cost management for development teams.
Developers increasingly rely on a diverse set of AI coding agents, such as Claude Code, Codex, and Cursor, for various tasks. However, this proliferation creates a significant governance challenge: each tool has its own command line, credential management, and method for executing shell commands, leading to a lack of clarity regarding agent actions and associated costs.
Omnigent emerges as an open-source solution to this problem, functioning as a meta-harness that sits above individual AI coding tools. It provides a unified interface for a wide range of agents, including Claude Code, Codex, Cursor, OpenCode, Hermes, Pi, and custom agents defined in YAML. Users can easily switch between or combine these agents with minimal configuration. Omnigent supports various authentication methods, drawing from a primary API key, a Claude or ChatGPT subscription, or compatible gateways like OpenRouter, with per-agent defaults that can be changed mid-session. Crucially, a single Omnigent session synchronizes messages, sub-agents, terminals, and files across different devices and platforms, including terminals, browsers, and phones.
A core security feature of Omnigent is its policy-driven governance. At the harness layer, administrators, developers, or individual users can define policies that dictate what an agent is permitted to do concerning shell commands, file edits, and token expenditure. These policies are enforced by checking every agent action, allowing it, blocking it, or requiring human approval. Built-in features include spend caps and access limits, enabling the establishment of hard dollar ceilings for sessions, complemented by softer warning mechanisms. By enforcing these rules at the meta-harness level, Omnigent ensures stateful, data-centric oversight that prevents agents from circumventing controls embedded within prompts.
Policies are structured hierarchically across three levels: server-wide rules set by administrators, per-agent rules defined by developers, and per-session rules established by the user. The system prioritizes stricter session rules. Omnigent also simplifies policy creation, allowing teams to write rules in plain language within a chat interface, which the agent then translates into executable policies.
To enhance security, each agent operates within an operating-system sandbox that restricts its access to the filesystem and network. On Linux, this isolation is achieved using bubblewrap for native terminal wrappers, while macOS employs the built-in seatbelt sandbox. These sandboxes can also hide credentials from the agent, allowing Omnigent to broker access and grant agents broad permissions while keeping sensitive secrets protected, a feature described as enabling "Run YOLO mode safely."
Windows support is currently offered in a degraded mode. Agents on Windows run under a Windows Job Object for process containment and resource limiting, but lack filesystem and network isolation. Windows users requiring full sandboxing are directed to use Linux, macOS, or the Windows Subsystem for Linux (WSL).
Collaboration is facilitated through shared sessions, where teammates can join by opening a link to observe an agent's activity in real-time. Features like history tracking, inline comments, and a side-by-side code editor enhance the collaborative experience. A "co-drive" mode allows a teammate to attach to a running session, executing their commands on the host's machine, which is ideal for pair programming or handing over control to a domain expert. This arrangement places responsibility on the host, making the policy layer and sandbox critical for maintaining control within shared sessions. Account access is invite-only, with options to integrate sign-in through Google, GitHub, Okta, or Microsoft.
The Omnigent repository includes two example agents: Polly, an orchestrator that plans and delegates tasks to sub-agents, assigns work to parallel git worktrees, and sends diffs for review by agents from different vendors before user merges; and Debby, a brainstorming partner that pits Claude and GPT models against each other in a debate before converging on a solution. Both examples highlight the framework's ability to leverage multiple AI models for a single task, enhancing the quality and diversity of output.
Omnigent can be deployed locally on a laptop or in disposable cloud sandboxes from providers like Modal, Daytona, E2B, CoreWeave, and Kubernetes. Sessions can be launched from the command line or provisioned per session, allowing machines to remain offline. The server can also provision a sandbox for each session, a setup termed "managed hosts," ensuring that work is carried out on an always-on host. Omnigent is freely available on GitHub.