# OmniAct

A lightweight, cross-environment general-action agent that completes real digital tasks spanning a browser, desktop applications and the command line, coordinated by a TeamLeader that dispatches to specialized Workers.

A TeamLeader decides whether to answer directly or hand a task to a Browser, CLI or Desktop Worker, each of which keeps the observation and action loop suited to its own environment. Workers hand off results as Markdown summaries and concrete artifacts rather than a bare completion message, written into a per-task directory alongside a trace log and a readable execution trajectory, so the collaboration stays inspectable.

Parallel subtasks, human takeover, resumable CLI sessions and execution traces support longer-running tasks; a Worker can pause for a human to take over and later resume. The CLI Worker drives Codex sessions, and the Browser Worker can reuse an existing Microsoft Edge profile in an isolated worker browser without taking over the user's own browser window.

It targets Windows with Microsoft Edge installed, needs Python 3.11 and OpenAI-compatible endpoints for the TeamLeader and the vision-capable Browser Worker, and is released under the MIT License.

Website: https://github.com/Iwtbs27/collab-agent