An open-source browser extension that puts AI agents directly into Chrome and Firefox sidebars has quietly processed more than a billion tokens monthly, its creator said, carving out a foothold among developers wary of handing browsing data to tech giants.
WebBrain, which launched publicly on April 28, 2026, offers something increasingly rare in the current AI boom: a path to automation that doesn't require funneling web activity through corporate servers. The free, MIT-licensed tool lets users run language models entirely on their own machines or connect to any cloud provider they choose, without intermediaries collecting data along the way.
That positioning matters. Anthropic, Google, and a handful of well-funded startups have spent the past year embedding AI assistants into browsers, but nearly all route user data through proprietary infrastructure. WebBrain sidesteps that model entirely.
"I built WebBrain because I wanted an AI browser assistant that didn't force me into one provider or require a subscription," said Emre Sokullu, the extension's creator, when he announced the launch in April.
The timing proved prescient, though whether by luck or foresight is hard to say. Corporate browser agents have matured quickly, yet concerns around privacy and vendor lock-in haven't disappeared—if anything, they've intensified as these tools handle more sensitive tasks.
How It Works
WebBrain appears as a sidebar panel capable of reading web pages, extracting structured data, and executing actions like clicks, form fills, and multi-step navigation. Two core modes govern its behavior. Ask provides read-only question answering. Act performs browser actions, but only after explicit user approval for each individual step.
A third Dev mode exposes page source code, network logs, and console output—tools aimed at developers debugging automated workflows. The agent can handle multi-step sequences with configurable limits, save workflows for reuse, and schedule tasks through a command-line interface.
The extension supports local inference engines including llama.cpp, Ollama, LM Studio, and vLLM, according to the project's GitHub repository. For those who prefer cloud processing, it integrates with OpenAI, Anthropic Claude, Google Gemini, Mistral, DeepSeek, xAI Grok, and over a dozen other providers through direct connections and OpenRouter.
WebBrain's documentation recommends pairing a fast text model for planning with a separate vision model to interpret screenshots. The extension compresses and resizes images to manage token consumption, a small detail that hints at careful resource management.
Security by Default
The extension runs in read-only mode unless a user grants broader permissions, surfacing capability-by-origin prompts before consequential actions. A "Strict Secret Mode" redacts credential fields before sending page content to language models—a guardrail against accidentally leaking passwords or API keys.
WebBrain's privacy policy, updated in early April, is unusually blunt: "We don't collect, store, or transmit any of your data." API keys live in the browser's local storage. Data flows directly from the extension to whichever model provider a user has configured, with no telemetry or analytics in between.
On Chrome, the extension uses the Chrome DevTools Protocol debugger API to generate trusted input events, which triggers a "WebBrain started debugging this browser" banner during active sessions. Firefox users get narrower permissions without debugger access—a limitation of that browser's architecture.
The Builder
Sokullu isn't new to building platforms. He founded Grou.ps, a Web 2.0-era social network that secured venture funding in 2009, and wrote for ReadWriteWeb during the early blogging era. His recent work has centered on open-source AI tools, though the trajectory from social platforms to browser automation isn't entirely linear.
Chris Messina, the hunter who featured WebBrain on Product Hunt, noted in a comment on the launch page that he appreciated "that @emresokullu's WebBrain extension gives people choice who use Chrome and other browsers." The phrasing was casual, but the sentiment reflected a broader frustration with walled-garden AI assistants.
Adoption and Usage
WebBrain surpassed 1 billion monthly tokens by July 26, 2026, and had accumulated over 500 GitHub stars as of that date, according to company press releases. The extension also recorded a single-day peak of 100 million tokens during July on its optional WebBrain Cloud service, which launched in June at $5 per month per device profile under a fair-use model.
Public metrics tell a more modest story. The Chrome Web Store lists 1,000 users and a 3.8-out-of-5 rating from 11 reviews. The Firefox Add-ons listing shows 84 users and a 5.0 rating from three reviews. Both stores carry version 30.0.3, the latest release.
An independent test published in July by Agent Workflow Lab described successfully running a local Qwen model via llama.cpp to fill out a form in WebBrain, stopping at a preview step without submitting. The researcher noted that the extension required explicit "Approve and execute" confirmations before each action and consumed roughly 53 gigabytes of VRAM—a steep hardware requirement that underscores the tradeoffs of local processing.
Beyond the Browser
WebBrain has begun extending into adjacent tools. An LM Studio plugin adds fetch_url and research_url functions and an optional bridge into a signed-in Chromium session, according to LM Studio Hub documentation. A Model Context Protocol server, launched via a Node.js command, lets coding assistants like Claude and Cursor delegate browser tasks to the live extension, though this feature works only on Chromium-based browsers.
A June update added a /record command on Chrome to capture interaction sequences and extended the scheduling window for automated tasks. The pace of development has remained steady, with blog posts in early August covering an MCP server guide and a local WebGPU vision fallback for text-only models.
The Landscape
WebBrain enters a crowded and fast-moving field. Anthropic piloted Claude integration in Chrome last year and expanded the feature to Pro subscribers in December. HARPA AI, a commercial Chrome extension with similar automation capabilities, reports more than 300,000 users. Sider AI markets cross-model support in a sidebar format across multiple browsers.
WebBrain's documentation includes a comparison section positioning the open-source extension against proprietary alternatives on three dimensions: local and offline support, multi-provider flexibility, and self-hosting capability. Those distinctions appeal to a specific audience—developers and privacy advocates who value control over convenience.
Whether that audience is large enough to sustain the project remains unclear. Open-source tools often struggle to compete with well-funded commercial alternatives that offer smoother onboarding and dedicated support. But WebBrain's growth, modest as it is, suggests there's appetite for an alternative.
Sokullu wrote on the Product Hunt launch page: "Today I'm launching WebBrain, a free, open-source AI browser agent that lives directly inside your browser." The project remains under active development, though how long one developer can sustain the pace against venture-backed competitors is an open question.
