Omnigent is a harness for your harnesses
The interesting AI-agent question in 2026 is no longer "which coding agent," it is "how do I run all of them." Omnigent answers that directly: it is an open-source meta-harness that puts a single orchestration layer over Claude Code, Codex, Cursor, OpenCode, Hermes, Pi, and agents you define yourself in YAML. You swap or combine harnesses without rewriting, and you drive them from a terminal, a browser, a phone, or a native macOS app, with sessions that stay in sync across all of them.
Install is one curl-piped script (or uv tool install omnigent), and omnigent picks a model with you and starts a session, also serving a local web UI at localhost:6767. It reuses credentials you already have: an API key, a Claude or ChatGPT subscription through the official CLIs, or any OpenAI- or Anthropic-compatible gateway like OpenRouter or Ollama.
Mixing vendors in one session
The headline capability is running different agents together and letting them check each other. The two bundled examples make it concrete. Polly is a multi-agent coding orchestrator that writes no code herself: she plans, delegates to coding sub-agents in parallel git worktrees, then routes each diff to a reviewer from a different vendor than the one that wrote it. Debby is a brainstorming partner with two heads, one Claude and one GPT, that answers every question twice and can debate itself before converging. Cross-vendor review and second opinions are exactly the kind of thing that is annoying to wire up by hand, and that is the gap Omnigent fills.
The operational story is serious for an alpha. You can run sessions in disposable cloud sandboxes from Modal, E2B, Daytona, Kubernetes, and others, so no laptop is required. Governance is first-class: policies can pause for your approval before risky actions, cap spend, or limit which tools an agent reaches, scoped to the whole server, one agent, or a single chat. On Linux and macOS the native terminal wrappers isolate each agent in a bwrap or seatbelt OS sandbox.
Where it fits, and the caveats
The honesty checks out in the docs. Omnigent is Apache-2.0 and explicitly alpha, and Windows runs in a "degraded mode": the SDK harnesses and web UI work, but the native terminal wrappers and the filesystem and network sandboxing do not, so Windows users lose the isolation that makes multi-agent runs safe to leave alone. Treat that as a real constraint, not a footnote.
If you are already juggling two or three coding agents and a pile of shell aliases to switch between them, Omnigent is worth installing just to feel what one governed, cross-device layer over all of them is like. Run it against a throwaway repo first, lean on the policies before you let anything loose, and judge it on whether the cross-vendor review actually catches things a single agent misses.
GitHub: https://github.com/omnigent-ai/omnigent
Curated by Agent Palisade — practical AI for small and mid-sized businesses.
Top comments (1)
The cross-vendor reviewer pattern is one of the more credible reasons to add orchestration above individual coding agents. A lot of teams talk about multi-agent setups, but having one harness plan, delegate, and then route review to a different vendor is a concrete control that can actually reduce same-model blind spots. I also like that you called out sandboxing and policy hooks early, because meta-harnesses get risky fast if they multiply tool access without adding boundaries. The next thing I would want in that stack is first-class trace visibility across agents so you can see where a bad handoff or bad review actually happened. How are you thinking about debugging a session once three different agents touched the same task?