Omnigent
Open-source platform to compose, govern, and collaborate on AI agents across multiple harnesses with built-in policies and sandboxing for secure execution.
Omnigent is an open-source alternative to:
Omnigent — A Meta-Harness for AI Agents
Omnigent is an open-source meta-harness that provides a common layer for composing, governing, and collaborating on AI agents—whether coding agents or others—on top of existing harnesses like Claude Code, Codex, or Pi. It enables seamless switching between harnesses, enforces stateful policies, and supports real-time collaboration across devices.
Key Features
- Built-in Agents: Includes Polly (coding orchestrator) and Debby (model debate) agents, with support to build custom agents in YAML.
- Contextual Policies: Enforce spend caps, model routing, and risk-based escalation with stateful, data-driven policies.
- Secure OS Sandboxing: Restrict filesystem and network access, hide credentials, and safely run agents in YOLO mode.
- Cross-Harness Composition: Combine multiple models and harnesses without rewriting code; switch between agents with one-line changes.
- Real-Time Collaboration: Share live agent sessions via URL with full history, enabling team review, inline comments, and joint steering.
- Multi-Platform Access: Interact via terminal, web UI, native desktop/mobile apps, or REST API.
- Extensible Architecture: Wrap any agent in a sandboxed session, add policies and shared history via a server layer.
Tech / Self-Hosting
Omnigent is open-source (Apache 2.0) and designed for self-hosting. It supports running agents locally or via cloud providers like Modal or Daytona. The architecture includes:
- Runner: Sandboxed execution environment for agents.
- Server: Adds policies, shared history, and exposes sessions across multiple interfaces.
- Clients: Terminal, web, native apps (macOS, iOS, Android), and REST API.
Built by the Databricks AI team and Neon, Omnigent emphasizes transparency, security, and developer collaboration.