Introducing Omnigent: A Meta-Harness to Combine, Control and Share Your Agents
Short summary
Omnigent adds a universal layer above existing agent harnesses—like Claude Code, Codex, and custom agents—to let you mix models, enforce stateful policies, and collaborate live in shared sessions. It provides one-line agent swaps, contextual security and cost controls, OS sandboxing, cloud execution, and real-time collaboration, all open-sourced under Apache 2.0.
- Databricks open-sourced Omnigent (Apache 2.0) as a meta-harness that sits above Claude Code, Codex, and custom agents, letting you swap models or harnesses with one line of code instead of rewriting logic.
- It enforces runtime policies rather than prompt-based guardrails—e.g., pausing a session after $100 in LLM spend or requiring human approval before a git push if new npm packages were installed.
- It adds real-time multi-user collaboration via shareable URLs, cross-platform clients (web, mobile, macOS), and cloud execution through providers like Modal, Daytona, Fly.io, or Railway.
- Agents are defined once in YAML and portable across harnesses, with an OS-level sandbox that intercepts network requests to control when secrets are injected.
Full summary
Databricks built Omnigent as a “meta-harness” that sits atop existing agent frameworks—Claude Code, Codex, Pi or custom SDKs—and turns them into interoperable components. Instead of juggling separate windows and copy-pasting between agents, Docs or Slack, you define agents once and swap models or harnesses with a single line of code. Omnigent layers in stateful policies to control costs, permissions and security at runtime, not with in-prompt guardrails. You can combine multiple models or subagents, track each session’s LLM spend (for instance, pause after $100), and enforce contextual rules like requiring human approval before a git push if new npm packages were installed.
On top of that common API, you get real-time collaboration via sharable URLs, a web/UI client plus mobile and macOS apps, and cloud execution on providers like Modal or Daytona. A flexible OS sandbox intercepts network requests—so you can inject secrets only when and where you approve. Authors define agents in YAML and port them across harnesses or mix them in one workflow without rewriting logic. The project is open-source under Apache 2.0, available now in alpha on GitHub, and already deploys to Fly.io, Railway or your own infrastructure. Future plans include meta-level optimization with GEPA, code-based introspection modules and broader harness integrations.