Waffy

Waffy is a free, open-source browser extension that puts an AI agent inside the tab you're already looking at. It works in two modes. In chat mode it reads the page content, summarises it, and answers questions without touching anything. In automate mode you hand it an objective and it takes over, clicking, typing, scrolling, and moving between tabs until the task is done. The distinction is deliberate, since most of the time you want an answer about a page rather than an agent loose on it. Keeping the read-only case separate from the acting case is a small design choice that turns out to matter a lot in daily use.

The design decision that defines the product is that you bring your own keys. Waffy supports OpenAI, Anthropic, Google, xAI, Groq, and OpenRouter, along with on-device models running in the browser, and the site is blunt that your keys stay in your browser. There's no account to create and no subscription. That flips the usual arrangement for AI browser tools, where you pay a monthly fee, your page content routes through a vendor's servers, and you take the privacy posture on faith. Here the extension talks directly to whichever provider you chose, and if you run an on-device model, nothing leaves the machine at all. For anyone whose browser is logged into work systems, that difference isn't academic, since the pages an assistant reads are often the ones you'd least like to hand to a third party.

The automation runs as a four-step cycle rather than one model improvising. It plans first, analysing the request and generating a structured plan. It executes, performing browser actions with live vision of the screen so it's reacting to what's actually rendered instead of a guess about the DOM. It validates, checking whether the task actually completed. Then it produces output, delivering a summary of what it did. That validation step is the one that separates useful browser agents from demos, because an agent that doesn't check its work will cheerfully report success on a form it never submitted.

Screen vision matters here more than it might sound. Plenty of automation breaks the moment a site ships a redesign, because the selectors it memorised no longer exist. An agent reading the rendered page has a better shot at surviving that, which is the difference between a script you maintain forever and one that mostly keeps working. It's not magic, and complex or unusual interfaces will still trip it, but the approach ages better than brittle selector chains. It also means the agent can work on sites you have no special integration with, which is most of the web.

The open-source claim holds up on inspection. The code is on GitHub under the Apache 2.0 license, written mostly in TypeScript, with the full source published rather than a marketing repository pointing at a binary. At the time of writing it had shipped a v1.0.2 release in July 2026 across roughly 167 commits, so it's under real development, though the star count is still small and the community around it is young. You can install a release or build from source with Node 18 or later and configure your own provider keys. Being able to read the code matters more than usual for a tool of this kind, since you're granting it the ability to act inside a browser that's signed into your accounts, and open source is the only way that promise is checkable.

It suits people who already pay for an AI provider and would rather not pay twice, plus anyone whose work involves repetitive browser tasks that never justified writing a proper script. Researchers pulling data off a series of pages, someone filling the same form across dozens of records, or anyone who wants a page summarised without pasting it into a chatbot. It lands well in that awkward middle ground where a task is too repetitive to enjoy but too irregular to be worth automating properly. The privacy-minded are an obvious fit too, since bringing your own key and optionally running locally removes the middleman entirely.

The honest limitations are the flip side of the same choices. Free and open source means nobody's on call for you, so support runs through GitHub and a Discord community rather than a helpdesk, and the site publishes no contact email. Bring-your-own-key means the extension costs nothing but your provider bill still runs, and agent loops that plan, look, act, and validate consume more tokens than a single chat message. Autonomous browser agents also remain an area to supervise, particularly anywhere logged in with real permissions. Chat mode's read-only behaviour is a sensible default for that reason. There's no pricing page to study, because there's no pricing.