browsirai. MCP server that connects AI agents to your running Chrome over CDP. not a headless browser. not a Playwright wrapper. your actual Chrome process, with all your cookies, sessions, and extensions intact. the agent never sees a login screen. you're already authenticated everywhere. one command: npx browsirai install

"show me my GitHub repos with star counts" every browser agent out there launches a clean headless instance. zero state. you either paste API tokens into the prompt or build OAuth flows. browsirai attaches to your running Chrome via CDP. your GitHub session cookie is already there. credentials never enter the LLM context window. same Chrome process you're using right now. same cookies. same localStorage. same service workers.

"block my calendar tonight 7-8pm for dinner" no Google OAuth app registration. no Calendar API credentials. no service account JSON. no permission scopes. it attaches to your Chrome where Google is already authenticated, navigates to Calendar, creates the event. every OAuth integration you've ever built was solving a problem your browser already solved.

"what am I working on right now?" browsirai reads your open tabs. 30 tabs of GitHub PRs, Notion docs, Grafana dashboards, Stack Overflow rabbit holes. other browser agents start with an empty page. browsirai starts with your actual browser state. your tabs are the highest-signal context source you have and no other agent can read them.

"check this repo for console errors and network requests" returns structured data. console messages with log levels, network requests with status codes and response bodies. all via CDP Runtime and Network domains. browser_snapshot returns accessibility tree text. ~500 tokens. browser_screenshot returns a base64 image. ~10,000 tokens. 20x token cost difference. most browser agents send screenshots. browsirai reads the DOM tree.

"which component renders this button? give me the source file." browsirai walks the React Fiber tree at runtime via CDP Runtime.callFunctionOn. parses jsxDEV() function bodies for fileName and lineNumber embedded by Babel's jsx-source transform. no source maps. no browser devtools extension. pure CDP. Element: <button> Component: CTASection Source: src/sections/CTASection.tsx:5:4 Component Stack: CTASection → App React, Vue (__vueParentComponent), Svelte (__svelte_meta). no other browser agent does Fiber tree walking.

This is exactly it. Protocol adoption wins when teams can enforce safe defaults at runtime. This is where long-term edge comes from.

thats really cool!

Thank you!!

Eline sağlık @mksglu dünden bu yana, Reddit thread'lerini taramak için kullanıyorum, çok temiz çözüm oldu.

Süper çok sevindim

amazing, keep going )

Using your actual Chrome session is the move. the auth problem kills most browser automation setups before they even start. I've been doing something similar with CDP for my own workflows, connecting ai agents to Edge over a tcp proxy in WSL2. the fact that cookies and sessions just work means you skip the entire oauth dance. one less thing to break at 3am when your scheduled tasks run. Curious how browsirai handles tab lifecycle. does it clean up tabs after the agent finishes or do you end up with 50 orphan tabs?

Noice work.

NICE!!!

using real Chrome with actual session state instead of headless is the right call half the reason automation breaks is because headless doesnt behave like the browser users actually run

this is the right approach. running against your actual browser session instead of headless means zero auth friction for any site you're already logged into. i've been using a similar cdp setup for automating workflows on authenticated sites and the difference vs playwright with fresh sessions is massive. no more juggling cookies or building oauth flows for every integration. the fact that this is an mcp server makes it even better since you can plug it into any agent that supports the protocol

elestirmek icin soylemiyorum ancak bir sorum var. bu cozume neden gerek var? "chrome in claude (beta)" veya diger browser use mcp'lerden ayristigi nokta nedir?

Selam Umut, eleştirsen de soru da sorsan ben yine mutlu olurum, birlikte geliştirelim. Haklısın, "chrome in claude" da aynı şekilde cookie/session kullanıyor. Temel fark şu: "chrome in claude" sadece claude'a bağlı, browsirai MCP standardı oldugu icin herhangi bir AI agent ile çalışabiliyor (cursor, windsurf vs). bir de developer tooling tarafı var, inspect_source gibi. onun dışında aynı mantık. Ben development yaparken kullanıyorum. Playwright ve Agent Browser headless browser olarak çalışıyor. Burada ise CDP kullanılıyor. CLI özelliği ile de QA-automation için bir aday.

Harika proje! Basarilar! Benzer islevi Playwright MCP extension + playwright-cli ile karsiliyorum. Ama bu yontem de gayet iyi. Ozellikle source inspection tarafinda guzel skill'lerle cok islevsel hale gelebilir.

connecting agents to your actual browser state is the missing piece. most mcp servers act like the browser is a fresh tab every time

Nice work on browsirai — the zero-auth angle is clever. If you ever want to charge per session or per agent call, does usage-based billing for MCP servers in 2 lines. Free to start, 4% on earnings only.