📡 MCP Radar

ig-shadow-walker/agentBrowser

A local browser your coding agent can actually finish a job in — file uploads, cross-origin iframes, native dialogs, and downloads, all via CLI/MCP.

4 stars
0 forks
TypeScript
momentum ▲ 8.0
created 2026-08-07
on radar since 2026-08-12
star trend 3 → 4 since 2026-08-12
ai-agentautomationbrowser-automationclaude-codecli-tooldevtools-protocolheadless-browserllm-toolsmacosmcpmcp-serverplaywright
View on GitHub ↗

About agentBrowser

A headless browser your coding agent can actually finish a job in.

Claude Code and Codex get a real Chromium they drive from the outside — with none of the restrictions browsers place on in-page automation. So they can log into your internal tools, fill the forms, upload the file, and be done.

We wanted Claude to do a genuinely boring job: log into our internal admin tool, find a record, and attach a document. It got about 80% of the way and then stopped dead — every single time, at the same step.

Browsers require a real human gesture before opening a file picker. That's not a bug or an oversight; it's a deliberate anti-abuse protection, and it's absolute. An agent working through a browser extension or injected JavaScript cannot attach a file to . No amount of cleverness gets around it, because the whole point of the restriction is that no amount of cleverness should.

Which meant our agent could do the interesting 80% and then hand the task back to a human for the tedious 20%. Worse than useless — it created work.

The fix turned out to be a change of vantage point. The restriction applies to code running inside the page. Drive Chromium from outside, over the DevTools Protocol, and setting a file on an input is an ordinary operation — no gesture required, because there's no page-level JavaScript asking for one.

That outside-the-page control comes from Playwright, which is what makes the whole thing possible. agentBrowser is the layer on top: an interface an agent can reason about, credentials it never gets to see, and a single binary you can hand to someone.

From the project README.

Maintaining this server?

Add the radar badge to your README — it shows your project was picked up by MCP Radar and links to this page:

[![On MCP Radar](https://mcp.liqiwa.com/badge.svg)](https://mcp.liqiwa.com/s/ig-shadow-walker--agentBrowser.html)

Related MCP servers

lordamdal/chromeboost

Give Claude hands in your own browser — a Chrome extension + MCP server that lets Claude Code drive your real, logged-in Chrome with human-shaped clicks, drags and keystrokes.

⭐ 4JavaScript
stanislawherjan1/gdocs-comments-mcp

MCP server for adding inline, range-anchored comments to Google Docs — the one comment operation the Google Docs & Drive APIs can't do. For AI agents (Claude, Cursor, Copilot) that review documents.

⭐ 3JavaScript
bolivestilo/Homekit

Homekit gives any AI agent direct, physical control over Apple Home. Flip lights. Lock doors. Read sensors. Through a single open interface.

⭐ 68TypeScript
scrollDynasty/multi-agent-browser-mcp

Multi-agent browser MCP runtime for isolated, persistent, parallel Playwright sessions. Run multiple AI agents without shared tabs or browser state.

⭐ 4TypeScript
ZoeLinUTS/MagicTeX-mcp

MagicTeX — a LaTeX editor for AI agents (Claude Code): live PDF preview, PDF comments that drive edits, Visual (WYSIWYG) mode, and git-tracked history. No local TeX install.

⭐ 3TypeScript
dhavanikgithub/chartbrew-mcp

Chartbrew + AI agents. MCP server exposing Chartbrew's documented API: teams, connections, datasets, dashboards, charts, live queries, and secure embedding. TypeScript · stdio · restricted/unrestricte

⭐ 3TypeScript

📬 Get the weekly radar in your inbox

The top new MCP servers of the week, every Monday. No spam, unsubscribe anytime.