Omnigent Open Source AI Agent Framework and Meta-Harness

Post Views: 5

Omnigent is an open-source AI agent framework and meta-harness designed to unify multiple coding agents into a single coordinated system.

Framework Overview

Developers frequently employ multiple coding agents for different tasks, utilizing tools such as Claude Code, Codex, or Cursor. Each platform operates with distinct command-line interfaces, credential handling methods, and approaches to executing shell commands within a working directory. This fragmentation creates challenges for teams in managing where agent activities occur and monitoring associated costs. Omnigent, an open-source initiative, functions as a meta-harness that integrates these tools, offering a unified layer for coordination.

Supporting Multiple Agents

Omnigent supports Claude Code, Codex, Cursor, OpenCode, Hermes, Pi, and custom agents defined in YAML format, allowing users to switch or combine them through simple configuration changes. The framework leverages a primary API key, a subscription to Claude or ChatGPT, or compatible gateways like OpenRouter, with per-agent defaults that can be adjusted during a session.

Session Continuity

A single session maintains continuity across terminals, browsers, and mobile devices, ensuring synchronization of messages, sub-agents, terminals, and files.

Security and Policy Enforcement

Security is reinforced through policy-driven controls. These policies dictate an agent’s capabilities regarding shell commands, file modifications, and token usage, with real-time checks to approve, block, or pause actions. Built-in features such as spending caps and access restrictions enable sessions to operate under strict financial limits, with intermediate alerts for approaching thresholds.

Policy enforcement occurs at the meta-harness level, utilizing stateful, data-centric rules that track agent behavior, avoiding reliance on prompts where deviations might occur. Three policy tiers exist: administrators define system-wide rules, developers set per-agent constraints, and session users establish per-session parameters, with stricter session-level policies taking precedence.

Sandboxing and Credential Management

Sandboxing and credential management are integral to the framework. Each agent operates within an operating-system-level sandbox, limiting filesystem and network access. On Linux, this isolation relies on bubblewrap, while macOS employs the seatbelt sandbox. Credentials are concealed from agents, with controlled access brokers enabling users to grant broad permissions while retaining secret management.

Platform-Specific Limitations

Windows support operates in a limited capacity, using Windows Job Objects for process containment and resource limits, though filesystem and network isolation is absent. Users seeking robust sandboxing are directed toward Linux, macOS, or WSL environments.

Collaboration Features

Collaboration is facilitated through shared sessions, enabling real-time interaction where team members can review history, add inline comments, and use a code editor simultaneously. A co-drive mode allows secondary users to attach to an active session, executing commands on the host machine to support collaborative workflows.

Deployment Flexibility

Deployment flexibility allows teams to run Omnigent on local machines or deploy sessions in disposable cloud sandboxes via providers like Modal, Daytona, E2B, CoreWeave, or Kubernetes. Sessions can be launched from the command line or provisioned dynamically, enabling offline operation. The server can also manage host environments per session, maintaining an always-on workspace.

Example Agents

The framework includes two example agents. Polly functions as a coding orchestrator, managing task planning and delegation by assigning work to parallel sub-agents within git worktrees. It routes code changes to reviewers from different vendors than the original authors, leaving merges to the user. Debby serves as a brainstorming partner with dual models—Claude and GPT—engaging in debates through a “debate command” before reaching consensus.

Accessibility and Availability

Access to sessions is restricted to invite-only accounts, with authentication options including Google, GitHub, Okta, or Microsoft. The project is freely available on GitHub.