Awesome AI Agents › Browser Automation Tools

microsoft/fara

⭐ 6251 Python added to this list on 2025-12-15 repository created 2025-10-29

Fara-7B is an efficient and compact agentic small language model developed by Microsoft specifically for computer use. With only 7 billion parameters, it is designed to operate as a Computer Use Agent (CUA) that interacts with computer interfaces such as mouse and keyboard to perform multi-step tasks on behalf of users. Unlike traditional chat models that generate text responses, Fara-7B perceives webpages visually and takes actions like scrolling, typing, and clicking on predicted coordinates without relying on accessibility trees or separate parsing models. This enables it to complete tasks efficiently with an average of about 16 steps per task, which is significantly fewer than comparable models. The model is trained using a novel synthetic data generation pipeline based on the Magentic-One multi-agent framework, utilizing 145,000 trajectories covering diverse websites, task types, and difficulty levels. Fara-7B is based on the Qwen2.5-VL-7B architecture and fine-tuned with supervised learning. It supports on-device deployment due to its compact size, which reduces latency and enhances privacy by keeping user data local. Fara-7B can automate a variety of everyday web tasks including searching for information, summarizing results, filling out forms, managing accounts, booking travel and entertainment tickets, shopping, comparing prices, and finding job postings or real estate listings. It achieves state-of-the-art performance on multiple web agent benchmarks, outperforming both similar-sized and larger models. Microsoft also released WebTailBench, a new benchmark with 609 real-world web tasks across 11 categories, where Fara-7B shows the highest success rates among computer-use models. The evaluation infrastructure includes Playwright for browser automation and an abstract web agent interface for model integration. Fara-7B is experimental and recommended to be run in sandboxed environments for safety. Overall, Fara-7B represents a significant advancement in efficient, agentic models for practical computer use, combining strong performance with compactness and privacy benefits.

https://github.com/microsoft/fara

agentagentic-small-language-modelbrowser-automationbrowser-usecomputer-usecomputer-use-agentcuafara-7blow-latencymagentic-onemicrosoftmouse-and-keyboard-interactionmulti-step-taskson-device-deploymentplaywrightprivacyqwen2.5-vl-7bsupervised-fine-tuningsynthetic-data-generationtask-automationvisual-operationweb-automationweb-benchmarkswebtailbench