Most teams end up juggling several agent harnesses, provider APIs, and ad-hoc safety checks. Omnigent adds a thin meta-layer above existing agents so you can treat different harnesses as interchangeable components, apply contextual policies across sessions, and share live agent workspaces for review and co-driving.
What Sets It Apart
- Uniform runtime across heterogeneous harnesses: wrap Claude Code, Codex, Cursor, Pi, OpenCode, Hermes, or your own SDK agents in a single session API so sub-agents can be mixed and swapped without rewriting prompts — this means you can split work across models and reviewers with one YAML change.
- Contextual, stateful policies and spend controls: policies inspect every tool call (shell, file writes, network, cost) and can allow, block, or pause for human approval; that lets teams enforce budget and safety rules at the meta-harness level rather than by brittle prompt engineering.
- Strong OS and egress sandboxing plus managed cloud execution: local bwrap/seatbelt sandboxes or disposable cloud sandboxes (Modal, Daytona, etc.) reduce risk when agents run code or network calls, while the server exposes sessions over terminal, web, desktop, or mobile for collaboration.
- Declarative agent authoring and composition: agents are compact YAML files (prompt, tools, sub-agents), making it straightforward to codify complex workflows, reviewers, and parallel sub-tasks.
Who It's For and Tradeoffs
Great fit if you operate multi-model agent workflows and need governance or team collaboration — e.g., engineering teams that want to run coding sub-agents, reviewers, and CI-like flows across vendor models. It centralizes auditing, spend caps, and human-in-the-loop approvals.
Look elsewhere if you only need a lightweight client or single-model chatbot: Omnigent is an orchestration layer with nontrivial dependencies and operational surface (Python 3.12+, sandboxing tools, optional cloud sandboxes). Windows support is limited compared with Linux/macOS, and the project is alpha-stage so expect rapid changes and integration work.
