479 ·GitHub
What is this?
OmniAct (repo name collab-agent) is a lightweight multi-agent system that helps you finish real digital tasks across browsers, desktop apps, and the command line.
OmniAct (repo name collab-agent) is a lightweight multi-agent system that helps you finish real digital tasks across browsers, desktop apps, and the command line.Is this a single agent or a team of agents?It's an orchestration system: a TeamLeader agent decides whether to handle a task itself or hand it to specialized Browser, CLI, or Desktop Worker agents, each keeping an observation-action loop suited to its own environment.
It's an orchestration system: a TeamLeader agent decides whether to handle a task itself or hand it to specialized Browser, CLI, or Desktop Worker agents, each keeping an observation-action loop suited to its own environment.How do I get it running?Set up a Conda/Python 3.11 environment, fill in your model service settings in the .env file, then launch the included script from the project root.
Set up a Conda/Python 3.11 environment, fill in your model service settings in the .env file, then launch the included script from the project root.Do I need any API keys or extra software before running it?Yes: a Windows machine with Microsoft Edge installed, Conda or another Python env manager, Python 3.11, an OpenAI-compatible endpoint for the TeamLeader's tool calling and the Browser Worker's vision, plus Codex CLI if you want to use the CLI Worker.
Yes: a Windows machine with Microsoft Edge installed, Conda or another Python env manager, Python 3.11, an OpenAI-compatible endpoint for the TeamLeader's tool calling and the Browser Worker's vision, plus Codex CLI if you want to use the CLI Worker.Is this ready to use as a finished product right now?Not really yet — it only runs on Windows with Microsoft Edge, and the public repo is still an active work in progress exposing just the core orchestration engine rather than a full product UI.
Not really yet — it only runs on Windows with Microsoft Edge, and the public repo is still an active work in progress exposing just the core orchestration engine rather than a full product UI.What does the TeamLeader agent actually do?It owns the overall task and only delegates clearly scoped subtasks to specialized Workers; for simple jobs it can also just reason and answer directly instead of splitting the work.
It owns the overall task and only delegates clearly scoped subtasks to specialized Workers; for simple jobs it can also just reason and answer directly instead of splitting the work.Can a human step in if a Worker gets stuck?Yes, Workers can pause and hand control to a person, and the project's roadmap specifically calls out improving recovery reliability after human takeover or partial failures.
Yes, Workers can pause and hand control to a person, and the project's roadmap specifically calls out improving recovery reliability after human takeover or partial failures.Where do I find the results of a task?Everything lands in a per-task folder under tasks/, including the task description, Worker requests and results, screenshots, file artifacts, and a readable execution trace.
Everything lands in a per-task folder under tasks/, including the task description, Worker requests and results, screenshots, file artifacts, and a readable execution trace.Why does the CLI Worker specifically need Codex CLI?The CLI Worker drives resumable command-line sessions through Codex CLI, which is the CLI backend the project currently integrates with.
The CLI Worker drives resumable command-line sessions through Codex CLI, which is the CLI backend the project currently integrates with.What are similar projects?magentic-ui, hive, midscene, typesafe-computer-use, Codex-Minecraft-Gameplay, oh-my-subagents, relay, agent-desktop
magentic-ui, hive, midscene, typesafe-computer-use, Codex-Minecraft-Gameplay, oh-my-subagents, relay, agent-desktop