My workflow looked like this: I'm building a webapp, I hit a bug. I open Chrome, inspect the page, take a screenshot. Switch to Claude, paste the screenshot, copy the console output, describe the context. Claude responds. I go back to Chrome to apply the suggestion. Repeat.
The worst part was when I needed to create an API key in some third-party tool. Claude would give me approximate paths. "Go to Settings > API > Create Key." Except the real path was three clicks somewhere else. Same thing for setting up ads, webhooks, anything that involved navigating a web interface.
The model was guessing. And I was the middleman between a blind AI and a browser that couldn't talk to it.
The actual problem
AI models work blind when it comes to the web. They don't have access to the page. Not the DOM, not the console, not the network. They mentally reconstruct what you describe, then propose something plausible.
Plausible, not correct. And the gap between the two costs you time.
What was missing wasn't a better model. It was a bridge between the model and the browser.
What I built
Eaight is a web browser (Chromium-based, built with Electron) that ships with a native MCP server. MCP is the protocol Claude Code uses to communicate with external tools.
In practice, it means Claude sees what you see. In real time. The page DOM, text content, accessibility tree, network requests, console logs, cookies. All exposed as MCP resources the model can read without you copying anything.
And it doesn't just watch. 18 automation tools give it hands: navigate, click, type, scroll, fill forms, execute JavaScript, take screenshots. Claude can go create its own API key in the interface, without me playing middleman.
How it connects
Claude Code detects the browser automatically. A discovery file, a WebSocket, that's it. No manual configuration, no token to paste. You launch Eaight, you launch Claude Code, they find each other.
For other CLIs (Codex, Gemini), a stdio bridge translates the requests. Same protocol, same result.
The timing
I started building Eaight in December 2025. A few months later, Anthropic released their own browser extension for Claude. Is it frustrating? A little. But I'm not the only developer on the planet. When a problem is real, multiple people see it at the same time. It just confirms the need was there.
And Eaight does things the extension doesn't: it's a full browser, not an extension. The MCP server runs inside the main process with direct access to tabs. No Chrome extension permission limitations.
What it changes
No more screenshots to paste. No more console to copy. No more invented paths. Claude looks at the page, understands the structure, and acts on it. The copy-paste workflow is gone.
The code is open source: github.com/monteiro-consulting/eaight
Top comments (0)