Ishmael

Agent Browser - MCP-powered browser automation for macOS & AI agents

by•
Do you have/want an Ai agent? Agent Browser is the ultimate ai agent browser for MacOS. Native with full GUI, and comes with proper multi tab browsing for you agent! Gives Claude, Open claw or any other AI agents a browser with all the tools they need to be fully autonomous locally on your Mac. Rated 9.5 by Claude! Full transcript of test done by claude online. Perfect for developers building AI agents, power users automating web workflows and also beginners just starting.

Add a comment

Replies

Best
Ishmael
Maker
📌
Hey Product Hunt! 👋 I'm Ishmael, the maker of Agent Browser. I built this because I kept running into the same wall: AI agents like Claude are incredibly powerful, but they couldn't actually *use* a browser the way a human would. Agent Browser solves that. It runs a real Chromium browser locally on your Mac, and exposes it to Claude and other AI models via the Model Context Protocol (MCP). Full stress test and review posted online.. Rated 9.5 by Claude! Your AI can now navigate pages, click, fill forms, extract content, and take screenshots as well as browse simultaneously on more than 1 tab! No cloud, no subscriptions, no data leaving your machine. Would love to hear your thoughts and answer any questions! 🚀
Julia Zakharova

@ishmael_dev Hi. MCP-powered browser automation for macOS - What about security here?

Ishmael
Maker

@julia_zakharova2 Great question, Julia! A few key points:

All traffic stays local. Both the browser daemon (port 8765) and MCP server (port 3100) bind exclusively to 127.0.0.1 nothing is exposed to the network or internet.

CORS is locked down. The HTTP server only accepts requests from the local dashboard origin (localhost:3210). The daemon itself has CORS disabled entirely.

Claude connects via stdio, not HTTP. The Claude Desktop integration uses a stdio transport, so there's no network socket involved on that path at all.

Session data stays on your machine. Cookies, localStorage snapshots, run history, and screenshots are stored in a local SQLite database.. never sent anywhere.

JS eval is a trusted-caller API. The evaluate action is explicitly documented in the code as for authorised agent use only not exposed to arbitrary user input.

The threat model is: a local tool controlled by you, talking to your own browser, with no cloud relay. Happy to dig into any specific concern!

Julia Zakharova

@ishmael_dev Thanks for the detailed answer