Playwriter

Most browser automation tooling spins up a clean, throwaway Chrome instance. That sounds tidy until an agent has to do anything real: it lands on a login wall, has no cookies, no saved sessions, no extensions, and is often flagged immediately by bot defenses. Cookie injection and credential replay are the usual workarounds, and they are fragile. Playwriter, from developer remorses, takes the opposite stance. It is a Chrome extension paired with a CLI and MCP server that lets AI agents drive the browser you are already using, with your logins, cookies, and extensions left intact. Instead of recreating your authenticated state somewhere else, the agent simply works inside it, which removes a whole category of brittle setup steps from agentic browser tasks.

The design has three layers. A Chrome extension attaches to a tab via the Chrome DevTools Protocol and opens a WebSocket connection to a local relay server on localhost:19988, which then exposes that connection to MCP clients and the CLI. Control is opt-in per tab: you click the extension icon, the icon turns green, and only that tab can be driven. What sets Playwriter apart from many MCP browser tools is the interface it gives agents. Rather than a fixed set of a dozen or so predefined actions, it exposes a single execute tool that accepts full Playwright snippets, so the model uses Playwright knowledge it already has and spends less context on tool definitions. It also favors compact accessibility snapshots over screenshots, plus network interception, live debugging, and stateful sessions across CLI calls.

The natural audience is anyone who wants an agent to operate inside a real, logged-in browser rather than a sterile sandbox. That covers developers wiring browser control into Claude Code, Cursor, or OpenCode, and people automating workflows on sites that are hostile to fresh headless sessions. Concrete uses include scraping or monitoring behind authentication without re-login, inspecting and reverse-engineering a site's API calls through network interception, setting breakpoints and live-editing page code while debugging, and capturing screen recordings of a flow. Because control is per-tab and a human stays in the loop, it also fits semi-supervised work where you let the agent run but step in to solve a CAPTCHA or clear a consent wall yourself, then hand control back.

The honest caveats are real. Playwriter is desktop-session-centric: it drives one person's running Chrome, not a fleet of server-side browsers, so it does not address large-scale parallel automation. It requires installing a Chrome extension, which is friction in locked-down or managed corporate environments where extension policy is restricted. The project reports a low fork count, which suggests adoption so far is more individual than enterprise, and that maturity gap is worth weighing. The README itself notes rough edges: a Chrome bug in the chrome.debugger API can leave pages returning about:blank until you restart Chrome, the browser may switch to light mode on connect, the automation infobar is left visible by design, and remote access requires explicit tunnel setup. Handing a capable agent full Playwright access to your authenticated browser is also a meaningful trust decision.

Against the commercial alternative Browserbase, the two solve different problems. Browserbase is managed, server-side browser infrastructure built for scale, zero-install provisioning, and concurrency, which is what you want when many agents need browsers on demand. Playwriter is deliberately the opposite: it lives on your machine, reuses your real session, and trades scale for authenticity and zero re-authentication. Adopt Playwriter when the point is to let an agent act inside your personal or work browser with all your existing context, or when you are building and debugging agentic flows against sites that resist fresh sessions. It is too early, or simply the wrong tool, when you need install-free, server-side automation across many sessions at scale. For a single trusted operator, though, it is a pragmatic and unusually capable option.