omnigent-reference

name	omnigent-reference
description	Standing pointer to the external `omnigent-ai/omnigent` project — MultiTable's closest prior-art / sibling. Trigger when the user mentions omnigent, omni, omnigent-ai, "compare to omnigent", "how does omnigent do X", `claude_native.py` / `codex_native.py` / `cursor_native.py` / `pi_native.py`, omnigent's `inner/` harnesses / `runner/` / `server/` / `policies/`, the `_HARNESS_MODULES` registry, `claude-sdk` vs `claude-native` harnesses, `is_native_harness`, the WS-tunnel transport, omnigent policy engine, omnigent governance, omnigent host registry / presence / attach / co-drive, omnigent cloud sandbox (Modal / Daytona / Islo), or wants a side-by-side of an omnigent concept against MultiTable.
allowed-tools	Read, Grep, Glob, Bash, WebFetch

omnigent-reference

omnigent-ai/omnigent is an open-source Python (~83%) + TypeScript (~16%) framework that solves a problem almost identical to MultiTable's: a unified harness over multiple AI coding agents (Claude Code, Codex, Cursor, Pi, Goose, Qwen, Kimi, Hermes, Copilot, Gemini/antigravity, OpenAI Agents, custom), with session sync, multi-agent supervision, governance policies, and cloud sandboxes. MultiTable is the user's variation of that same problem, written in Node/TS.

Architecture headline (read before the reference files): omnigent runs every provider through two parallel layers, registered by name in _HARNESS_MODULES (omnigent/runtime/harnesses/__init__.py):

1. Inner harnesses (omnigent/inner/<vendor>_harness.py) — headless/SDK/ACP/--print execution (e.g. claude-sdk, codex, goose acp, , , , ). These are the analogue to MultiTable's adapters.