WebVoyager

Category: OSS Agent

License: Apache License 2.0 (free and open source)

Official repo: https://github.com/MinorJerry/WebVoyager

Research paper: arXiv:2401.13919

---

What it is

WebVoyager is an end-to-end web agent powered by a Large Multimodal Model (LMM). Given a natural-language task, it autonomously navigates real-world websites by combining visual perception (screenshots) with textual understanding to click, type, scroll, and complete multi-step instructions without human intervention.

It was built by researchers at Zhejiang University and Tencent, and is not an officially supported Tencent product. The authors disclaim responsibility for content output by the model, any web pages visited, or data obtained during operation.

Key capabilities

• Completes user tasks end-to-end on live websites
• Combines visual (screenshot) and textual (accessibility tree) input modes
• Supports headless Chrome or visible browser mode
• Ships with 643 benchmark tasks across 15 websites (data/WebVoyager_data.jsonl)
• Includes an automated GPT-4V evaluation pipeline

Cost model

The tool itself is free and open source. Running it requires an OpenAI API key (GPT-4V by default for vision mode, GPT-4 Turbo for text-only mode). API usage is billed by OpenAI at standard rates.

Safety and ToS note

WebVoyager operates a real Chrome browser and submits real requests to live websites. Automated browsing may violate the terms of service of some websites. Review target site ToS before running production workloads. The repo disclaimer explicitly states the authors bear no responsibility for model output, web pages viewed, or data obtained.