BrowserKitten is an MIT-licensed Chrome Manifest V3 extension that puts an agent in the browser side panel, for people who want to read, act on and produce files from the signed-in pages they already use, driven by models they configure themselves.

What it is

BrowserKitten is a JavaScript Chrome extension built on Manifest V3, with the agent living in the side panel rather than in a separate application window. It works with the browser session the user is already signed in to and keeps conversations, workspace files and follow-up tasks on that computer, while model and other provider requests go to the services the user configures. The current source version is 1.2.0, and its topics include ai-agent, browser-agent, byok, local-first and manifest-v3; it sits in the Data & Analytics / Data Extraction & Web Scraping category.

The concrete problem is the gap between an authenticated page and the tool meant to work on it. Agents driving a fresh browser profile cannot reach content behind a login, so the usual fallback is to copy text out by hand or to maintain separate automation software. BrowserKitten reads the pages the user is signed in to, selects relevant content and collects findings in place, and it does so under a bring-your-own-key arrangement: the unpacked build accepts the user's own keys without a Browser Kitten subscription, and account-hosted services are optional and depend on a separate backend.

Key capabilities

• Reads and researches signed-in pages, selecting relevant content and collecting findings.
• Delegates clicks, form entry and navigation to a separately configured execution model, System1.
• Creates deliverables such as documents, HTML pages, data and files in a session workspace.
• Runs local follow-ups from a collapsible task drawer above the input, where the user states the goal, permitted actions, reporting conditions and stop conditions, and the agent arranges or removes wakeups within that goal.
• Configures Inference (base URL, API key, tool-capable model ID) and System1 (an OpenAI-compatible /chat/completions endpoint) separately, though both may use one provider.
• Adds optional Image, Transcription and Search services when a task needs them.
• Installs with no build step: the folder containing manifest.json is the load root for Load unpacked, and Node.js, Python and npm are not required.

Who uses it and how

• Researchers and analysts working behind a login, who summarise the current page and save its key findings as a document without leaving the browser.
• Users delegating page interaction to an execution model kept distinct from the planning model used for analysis and tool calls.
• Teams with own-key or data-ownership requirements: conversations, workspace files and follow-up tasks stay on the machine, and requests reach only the providers they configure.
• People watching a page over time, who set a follow-up goal and receive reports on actions, changes or requests for help, while quiet scheduled tasks stay silent when nothing meaningful changes.

Getting started

Download the current source ZIP from the main branch, extract it, open chrome://extensions with Developer mode enabled, and use Load unpacked to select BrowserKitten-main, the folder that directly contains manifest.json. Pin the extension, open the side panel and configure providers in Settings; a small first task such as summarising the current page and saving its key findings as a document is a reasonable check.

How it compares

No comparable tool is named in the facts for this registry entry, so BrowserKitten stands alone here. Its position is set by MIT-licensed source, an unpacked install with no build step, bring-your-own-key access to OpenAI-compatible endpoints, and conversations, workspace files and follow-up tasks held on the local computer.

When to use it — and when not to

Choose it if you want an agent inside the browser profile you are already signed in to, you accept configuring and paying for Inference and System1 separately, and you can leave Chrome running and the computer awake so minute-scale follow-up checks fire. Do not choose it if you need cloud execution, native OS or terminal control, immediate webpage-event subscriptions, or automatic grouping of every agent-controlled tab, all of which the README lists as current limits. Expect to validate behaviour yourself: forks are few against the star count, and the README points users to main rather than to tagged releases.