AI Browser Agent

AI Browser Agent gives your browser to an AI model you choose, and keeps the safety controls out of that model's hands. You supply the model — an OpenAI-compatible endpoint, Anthropic, or Gemini, using your own API key. The extension supplies everything else: reading pages, filling forms, clicking, navigating, working across tabs, inspecting the console and network when something is broken, and asking you before anything consequential happens. WHAT IT CAN DO • Read a page as structure, not as a screenshot — roles, labels and states, so the model works with what a screen reader would see. • Fill forms: text, dropdowns, checkboxes, radios, multi-selects, and structured inputs like dates, times, colours and ranges. • Click, scroll, wait, and navigate, re-checking the page each time. • Work across several tabs, with each task owning the tabs it opened. • Diagnose a broken page by reading its console and network activity. • Upload a file you pick yourself, and download one if you allow it. • Record a task you performed and replay it later. • Run bundled skills, and name them as shortcuts. • Connect GitHub to read and write issues, or Figma, Jira and Confluence to read from them, using a token you create in your own account. • Keep an audit trail of every decision, which you can export. HOW IT DECIDES WHAT IT MAY DO Not by asking the model nicely. Every action is rated for risk and checked against a policy the model cannot reach or change. Anything consequential asks you first. Some things are refused in every mode, including payments, entering card or identity details, submitting credentials into a page that did not issue them, permanently deleting records, and defeating CAPTCHA. If a page tries to hijack the agent — and pages do try — page content is wrapped as data that cannot close its own envelope and become instructions. The policy engine runs on what the browser actually did, not on what the model was told. Data that came from one site does not silently travel to another. The extension tracks which sites a task has read from, and a request carrying that data to somewhere else is blocked or asks you first. Credential-shaped values are removed at the moment they are collected, before anything stores or displays them. WHAT YOU NEED An API key for one of: any OpenAI-compatible endpoint, Anthropic, or Google Gemini. The extension ships no key and no free tier, and sends your key only to the provider you chose. The model must support tool calling. The extension checks this when you connect and tells you plainly if a model cannot drive a browser, rather than starting a task that will fail halfway through. WHAT IT DOES NOT DO YET Being straight about this up front, because discovering it after installing is worse: • Four connectors exist: GitHub, Figma, Jira and Confluence. There is no Google Sheets integration. Figma, Jira and Confluence are read-only, because none of those services tells the extension what a token may do — so rather than attempting a change and failing, it offers no way to make one. • You connect each one with a token you create in your own account, not by signing in. Jira and Confluence also need your own site address, and your token is bound to that one address and can go nowhere else. • No connector has been used against a real service from this repository. The paths are tested against local servers; what is untested is somebody else's production endpoint. • Plugins are not implemented. • MCP works as a client only. You can add a server in Settings and its tools and resources are offered to the model, confirmed on every call. It has been verified against the MCP project's own reference server; no server that somebody else operates has been reached from here, and the extension is never itself an MCP server. • Scheduled tasks run only while Chrome is running. There is no cloud scheduler, so a schedule due while the browser is closed is skipped rather than caught up, and a scheduled run stops instead of approving anything on your behalf. • It does not work inside cross-origin iframes — a form in an embedded frame is out of reach, deliberately, because reaching into every frame on every page is a much larger risk than the feature is worth. PRIVACY No telemetry. No analytics. No error reporting. Your data goes to the AI provider you configured, to a connector you connected, and to the page you are working on — nowhere else. It does not read cookies, session tokens, passwords, browsing history or bookmarks, and does not request the permissions that would let it. It cannot browse your computer: there is no filesystem access, and a file reaches the agent only when you pick it in Chrome's own file picker. Everything it stores stays in your browser.