This week the agent moved into the browser.

A JavaScript library from Alibaba (MIT, 19.6k stars, v1.10.0 shipped June 15) that turns any web app into something an agent controls in plain language — from inside the page. Drop in a <script> tag or npm install page-agent, point it at any OpenAI-compatible endpoint with your own key, and agent.execute('Click the login button') works by reading and manipulating the DOM directly. No screenshots, no headless Chrome, no extension, no Python sidecar.

Reach for it when you want an agent to operate your product's interface — an in-app copilot, a "do this for me" button — without standing up a vision model and a remote browser farm. It replaces the screenshot-and-click computer-use stack for the case where you already control the page. The caveat: it drives the DOM, so it's only as good as your markup — canvas-heavy or aggressively obfuscated UIs give it less to grab — and at 19.6k stars this is an established project riding a fresh release, not a debut.

A Chrome/Firefox extension (Apache-2.0, three days old at 0.x) that runs the entire agent loop client-side: it reads your tabs, decides, acts, and verifies against the live page, with no backend and no telemetry. The interesting part is the compute it carries — a CheerpX-emulated Debian Linux VM in WebAssembly with a real bash shell, sealed Web Worker notebooks for JS, and sandboxed iframes the agent can build apps into. Keys (Anthropic, OpenRouter, or local Ollama) sit encrypted in a local vault, and the agent that holds them never touches a raw page itself — a disposable runner does, then hands back fenced output.

Reach for it if you want an autonomous browser agent but won't ship your cookies and API keys to someone's server to get one. It fills the gap between "hosted agent that sees everything" and "no agent at all." The honest caveat: it's a 0.x experimental beta you load unpacked from the source tree (store approval pending), the peer-to-peer agent-to-agent layer is preview-only, and at 91 stars this is a bet on an idea, not a dependency you pin.

A local CLI that runs alongside your native Chrome so an agent can act on sites you're already signed into — your cookies, sessions, and 2FA state stay on your device. It's built for the place most automation dies: login walls, CAPTCHAs, anti-bot checks. It keeps multiple isolated sessions, pauses for a human to clear a challenge, then resumes from the same state, and returns cleaned structured data instead of raw HTML.

Reach for it when the agent's job is on the authenticated web — your dashboards, your vendor portals — and a fresh headless browser just hits a login page. It replaces the brittle "give the bot a service account and a stale cookie jar" pattern with your own live session, on your machine. Tradeoffs: it's free-for-now behind a 7-day-trial framing, so pricing isn't settled, and pointing an agent at sites with your real credentials is exactly the access you should think hard about before automating.

A CLI (Apache-2.0, npm install -g browserbash-cli) that turns natural-language descriptions into browser tests and drives a real browser to run them — no selectors, no locators, no Playwright boilerplate. It runs against local Chrome or any CDP endpoint, and it'll run on free local models (Ollama) or a free OpenRouter key, so the whole loop can stay on your machine at zero marginal cost.

Reach for it when your end-to-end suite is a graveyard of flaky CSS selectors and you'd rather describe the flow than maintain the locators. It replaces hand-written Playwright or Cypress specs for the happy-path coverage that breaks every time the markup shifts. The tradeoff: natural-language-to-test is only as repeatable as the model behind it — there's a paid cloud dashboard for run history and video past 15 days, and you'll want to pin the behavior you actually care about rather than trust a re-interpreted prompt every run.