If you use Claude Code for front-end or full-stack work, you've probably hit the same wall I did. The agent can read your code and run your tests, but it can't see the page. When it can, it's usually a fresh headless browser: no cookies, no logins, no staging SSO, and none of the state you were looking at when the bug happened.

Nav8 takes the opposite approach. It's a Chromium-based (CEF) browser for Apple Silicon Macs, and it exposes itself, your real browser with your real sessions, to Claude Code over MCP.

Setup: there isn't really any

On launch, Nav8 registers itself as a stdio MCP server in ~/.claude.json (it saves a backup to ~/.claude.json.nav8-bak first). Then:

claude

The nav8-browser tools are already there, in any terminal. There's no extension to install and no separate Chrome profile to keep logged in.

What Claude Code gets

The tools fall into two groups.

DevTools-grade inspection

• Console messages
• Full-fidelity HAR, captured through the CDP Network domain
• Performance profile and trace
• Core Web Vitals
• Device emulation

Driving the page

• Click / fill / type
• PageMap: a structural map of the page (forms, buttons, links, data) instead of a screenshot
• Parallel multi-tab work
• Schema extraction
• Cookie-authenticated fetch (requests go out with the session you're already signed into)
• JS eval and screenshots

Because all of this runs in your actual browser, "check the network calls on the billing page of staging" works even when staging sits behind a login. You're already signed in.

"Highlight for Claude"

This is my favourite part for UI work. In Nav8 you click an element on the page and attach a note ("this button wraps on narrow screens"), and Claude Code receives the exact DOM node along with your note. You don't have to describe which div you mean or paste selectors. You point, and the agent knows.

A typical loop

1. Ask Claude Code to fix a layout or performance issue.
2. It opens the page in your Nav8 tab, reads the PageMap, console and HAR, and checks Core Web Vitals.
3. It edits the code, reloads, and measures again.
4. You highlight anything that still looks wrong, and it gets the exact elements.

Every click happens in a real tab you can watch, and you can step in at any point.

It's an MCP host too

It works the other way as well. Inside Nav8 you can add any MCP server (a local stdio command or a hosted Streamable-HTTP URL) and scope it to specific agents. New servers get the narrowest scope by default. MCP secrets are stored in the macOS Keychain and injected at spawn, so .mcp.json only holds a ${NAV8_MCP_…} placeholder.

Safety rails

The agent is free to read, search, compare and draft. Anything irreversible stops on an approval card first: buying or paying, sending an email or message, deleting, or granting OAuth/camera/mic access. Agents run as a local process on your Mac, and model costs go to your own provider account (it works with the Claude or Codex subscription you already have).

Try it

Nav8 is free during the private beta. It runs on Apple Silicon Macs (M1 and later). Access opens in batches: you leave your email and get a download link when your batch opens.

Developer docs and beta signup: https://nav8.ai/developers?utm_source=hashnode&utm_medium=blog&utm_campaign=dev-mcp

I'd love to hear how you'd use a signed-in browser from Claude Code, and what tools you'd want added next.