BrowserKitten is a Chrome extension that turns the already-logged-in browser into a web agent under an MIT license, and picked up 2,906 stars in 25 days: Reveneau

Why it mattersA production-quality browser agent that ships as a plain Chrome extension and reuses the user's own session is the shape that gets past the retailer blocks and single-sign-on walls a cloud-hosted agent runs into, so a team building agentic tooling for logged-in workflows can now hold up a real MIT-licensed reference to design against.

Player-YN released BrowserKitten, a Chrome Manifest V3 extension that turns the already-logged-in browser into a programmable computer for an AI agent, on 28 August 2026. The repository, at github.com/Player-YN/BrowserKitten, reached 2,906 stars and 10 forks in the 25 days since, under an MIT license. The project name inside the repository is Paw Work.

What it does

The extension gives an agent access to the current tab: it can drive the DOM, run JavaScript on the page and keep artifacts, meaning spreadsheets and documents produced during the session. The workflow is selection-first, described in the README as "select on the live page, describe the outcome, take away an editable office file". Nothing runs on a server. The model call is bring-your-own-key against an OpenAI-compatible endpoint, and the key sits in chrome.storage.local on the user's machine.

The traction is real

2,906 stars in 25 days, with 10 forks, is a healthy ratio. A star-to-fork ratio like that reads as reader interest rather than the fork-farm pattern common on inflated repositories, where forks meet or exceed stars. GitHub's own API confirms the numbers and the 2026-08-28 creation date. The most recent push is 18 September 2026, so the repository is being maintained.

Why the shape matters

A cloud-hosted web agent visits a site from outside the user's session. To reach an account it has to hold that user's credentials on its own side, and the site receives a request from a server rather than from the person who holds the account. A site inspecting its own traffic can see both of those.

BrowserKitten is the opposite architecture. The agent runs inside the user's own Chrome, from within the tab the user opened, using the session cookies the user's login already produced, and the model key never leaves the machine. The site receives the user's own request, because it is the user's own request. That does not make the design legal against every site's terms, and it does not answer the question of what an agent should be allowed to do with an account once it is inside one.

What is on the page and what is not

The repository is fully open source and MIT-licensed. The extension has to be loaded unpacked from a clone; it is not on the Chrome Web Store. The README lists artifact management, session-first design and BYOK as the differentiators against hosted web-agent tools. Which specific OpenAI-compatible models are supported beyond the interface is not stated on the page, and no benchmark or demo video is linked. Anyone considering the design as a reference should read the AGENTS.md files in the repository, which is where the architecture is documented rather than in the README.

For a team building agentic tooling for logged-in workflows, the trade is worth naming out loud: only one user is served at a time, and every credential the agent can touch is a credential the user already has in Chrome. That is a smaller list of accounts to worry about than a server-side agent holds, at the cost of not scaling to many users on the same account.

Source

GitHub, Player-YN/BrowserKitten, MIT license, verified by the GitHub API at 2026-09-21T17:35:00Z.

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.