#4848th of 54 in Agent frameworks and SDKs
agent-s
Computer-use agent that operates desktop GUIs via screenshots, mouse and keyboard
Live demo ↗
(opens in a new tab)Website ↗
(opens in a new tab)Repository on GitHub ↗
(opens in a new tab)
- Stars
- 12.7k
- License
- Apache-2.0
- Last commit
- Oct 2026
- Last release
- Dec 2025
- Language
- Python
10,525 on 31 Mar 202612,666 on 20 Oct 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
Agent S takes a natural-language task, reads the screen, and completes it by clicking, typing and scrolling in ordinary desktop and web apps on macOS, Windows and Linux. It pairs a main LLM (OpenAI, Anthropic, Gemini, Open Router, vLLM, Azure) with a grounding model such as UI-TARS-1.5-7B, and ships as the gui-agents Python package with a CLI. Agent S4 is a separate package in the same repo.
Who it is for: Researchers and developers building or testing desktop computer-use agents
Strengths
- Works on any GUI without per-app API integration or scripting
- Runs on macOS, Windows and Linux; pip install gui-agents plus a CLI
- Reports 72.60% on OSWorld using multi-rollout selection; papers are published
- Main and grounding models are configurable, including vLLM and custom endpoints
Weaknesses
- Requires a separately hosted grounding model such as UI-TARS-1.5-7B
- Executes generated Python on your machine; no built-in sandbox
- Designed for single-monitor setups only
- Recommended setup uses paid OpenAI models; hosted Sai is the production offering
What it needs
- GPU optional
- Needs Tesseract (pytesseract), Grounding model endpoint (UI-TARS-1.5-7B), LLM API key (OpenAI, Anthropic, etc.)
- Models: OpenAI, Azure OpenAI, Anthropic, Gemini, Open Router
Also in Agent frameworks and SDKs
See all 54| Rank | Project | Score |
|---|---|---|
| 1 | ppt-masterAgent skill that turns documents into natively editable PowerPoint decks | 94 out of 100 |
| 2 | langchainPython framework for building LLM agents and apps from interoperable components | 89 out of 100 |
| 3 | browser-usePython library that lets an LLM agent control a web browser | 89 out of 100 |
| #4 | tradingagentsLangGraph multi-agent framework simulating a trading firm to analyze tickers | 88 out of 100 |
| #5 | copilotkitSDK for adding agent chat, generative UI and shared state to apps | 87 out of 100 |