- Published on
An Open-Source, Local-First AI Browser Agent That Reads Pages and Automates Tasks in Chrome and Firefox
- Authors
- Name
- DeepHorus
A new open-source browser agent called WebBrain has been released for Chrome and Firefox, allowing users to automate tasks, extract data, and interact with web pages using natural language. Developed by Emre Sokullu and licensed under MIT, WebBrain operates either entirely locally or via cloud APIs, offering a privacy-first approach for AI-powered browsing. Unlike many browser plugins that send data to external servers, WebBrain can run on a local model, ensuring no page data leaves the user's machine. This makes it a compelling option for privacy-conscious users and organizations handling sensitive information.
WebBrain integrates into the browser's side panel, with separate conversation histories per tab. It supports two modes: Ask mode, which is read-only and can summarize or extract data without altering the page, and Act mode, which can click, type, scroll, and navigate. Act mode leverages the Chrome DevTools Protocol, producing trusted input events that work across cross-origin iframes and shadow DOM. However, Firefox lacks CDP support, making its Act mode significantly weaker. The extension ships in five languages including English, Spanish, French, Turkish, and Chinese, with automatic language detection on first launch.
Key security features address prompt injection risks inherent in browser agents. WebBrain starts in read-only Ask mode and prompts users before any consequential actions, though these prompts can be disabled in settings. A UI-first rule prevents the agent from directly calling REST or GraphQL endpoints for mutations like form submissions or purchases, requiring visible UI interaction. A per-conversation override exists for cases where the UI fails. Data extraction and research tasks use background HTTP via the fetch_url and research_url tools, as they are read-only and pose no mutation risk.
Practical use cases include data extraction from product catalogs or PDFs, research summaries that respect paywalls, form filling with an optional local profile store, and multi-step automation like navigating to GitHub and finding trending repositories. The agent also dismisses cookie-consent banners before reading, streamlining the experience. Token costs are managed through resized and compressed screenshots, oldest-first trimming of conversation history, and the ability to pair a cheap text model with a separate vision model.
WebBrain positions itself between simple browser AI plugins and full agent frameworks. Its open-source nature and MIT license encourage community contributions and audits. The developer, Emre Sokullu, has not provided official reactions, but the GitHub repository indicates active maintenance and a welcoming stance toward contributions. The extension currently supports Chrome with Manifest V3 and Firefox with Manifest V2, reflecting ongoing browser ecosystem transitions.
Looking ahead, WebBrain's local-first design could influence the development of privacy-preserving AI tools. As browser agents become more common, the security model presented here—with explicit prompts and UI-first mutation rules—may set a benchmark. Continued updates are expected to refine Act mode on Firefox and expand language support. For now, WebBrain offers a powerful, transparent way to automate browsing tasks while keeping data under user control.
