Claude in Chrome
What if your coding agent could see exactly what you see in the browser — logged in, real data, live state?
Claude in Chrome is Anthropic's browser extension that connects Claude Code to your Chrome session via MCP. The agent navigates pages, clicks elements, fills forms, reads console logs, monitors network requests, and records GIF captures — all through your authenticated browser session.
How It Works
The extension runs as an MCP server inside Chrome. Claude Code connects to it and gets access to tools:
| Tool | What It Does |
|---|---|
navigate | Go to URL, wait for load |
read_page | Screenshot + DOM snapshot of current tab |
computer | Click, type, scroll at element level |
form_input | Fill form fields by label or selector |
javascript_tool | Execute arbitrary JS in page context |
read_console_messages | Read browser console with regex filtering |
read_network_requests | Monitor XHR/fetch traffic |
gif_creator | Record multi-step interactions as GIF |
tabs_context_mcp | List and switch between tabs |
The key difference: it operates inside your real browser session. No headless environment, no auth setup. Whatever you're logged into, the agent sees too.
Strengths
| Area | Detail |
|---|---|
| Session reuse | Zero auth config — inherits your cookies, tokens, everything. Richest authenticated access. |
| Action breadth | Deepest action model — navigation, forms, JS execution, console, network, screenshots, GIFs. |
| Debug power | Console log filtering + network monitoring = agent can debug its own frontend work. |
| IDE integration | Works from Claude Code CLI and VS Code extension. |
| Human-in-loop | Tab isolation, confirmation prompts for destructive actions. |
Limitations
| Area | Detail |
|---|---|
| Vendor lock | Claude only. Cannot use with Gemini, Codex, Cursor, or other agents. |
| Browser lock | Chrome and Edge only. No Firefox, Brave, Arc, Safari. |
| Token cost | Screenshots are expensive — 10-20K tokens per page read. Adds up in long sessions. |
| Extension fragility | Extension updates, Chrome updates, or MCP connection drops can break mid-session. |
| Proprietary | No standard body. Anthropic can change the API at any time. |
Setup
- Install the Claude in Chrome extension
- In Claude Code, run
/mcpand connect toclaude-in-chrome - Call
tabs_context_mcpto see your current browser state
Working demo in under 10 minutes. The extension handles all the MCP plumbing.
Best For
- Developer workflow automation — "build this component, open it in the browser, verify it renders, check console for errors"
- Testing authenticated flows without headless auth setup
- Debugging frontend issues with console/network inspection
- Recording browser interactions as GIFs for documentation or review
Checklist Score
Against the decision checklist:
| Gate | Score | Notes |
|---|---|---|
| 1. Perception | Good | Full page read with screenshots, DOM, and console. Rich but token-heavy. |
| 2. Action | Strong | Deepest action model — navigation, forms, JS, console, network, GIFs |
| 3. Auth | Strong | Best in class — full session reuse, zero config |
| 4. Workflow | Moderate | Claude Code only. No CI story. VS Code integration is a plus. |
| 5. Architecture | Strong | External tool — zero app changes |
| 6. Standards | Weak | Single vendor, proprietary, Chrome/Edge only |
| 7. Speed/Cost | Weak | Extension overhead + screenshots = most expensive per interaction |
| 8. Setup | Good | Extension install + MCP connect. ~10 min. Can be fragile across updates. |
| 9. Blast Radius | Strong | Tab isolation, human-in-the-loop, scoped permissions |
Context
- AI Browser Tools — Decision checklist and radar
- Claude Code — The CLI that connects to this extension
- AI Coding Config — Multi-agent setup