← Blog

How to give Claude Code access to your browser

2026-06-25

If you build with Claude Code, you already know the wall: the agent is context-blind to your browser. It writes code well, but the moment something is wrong in the running app — a layout that's off, a flow that breaks, a request that 401s — you're back to typing paragraphs to explain what you saw, or pasting screenshots and stack traces and hoping the agent guesses right.

ThinkRun closes that gap two ways: an MCP server that drives your real Chrome session, and a recorder that turns "what happened" into a structured artifact the agent can read.

The fastest path: the ThinkRun MCP server

Most browser MCP servers spin up a fresh, headless Chromium — so they lose your cookies, get blocked as bots, and can't touch anything behind a login. ThinkRun's MCP uses your real, logged-in browser, so Claude Code can act on authenticated pages the way you would.

npm install -g @thinkrun/cli
thinkrun setup        # installs the native-host bridge — no API key for local mode

Then add the MCP server to Claude Code and ask it to navigate, click, extract, or screenshot the page it's working on. It's the same browser you're already signed into.

When you just need to show the agent what's wrong

For a bug you can see but can't easily describe, record it instead of typing it:

  1. Hit record, reproduce the problem in your browser.
  2. ThinkRun captures the video plus the structured signal — the actions you took, DOM snapshots, console errors, and network requests — and runs an analysis pass.
  3. Share the result. The agent reads a .md or .json bundle (not a wall of prose): what happened, the repro steps, expected-vs-actual, and the first bad signal.

That's the difference between "the agent optimizes a symptom for an hour" and "the agent fixes the root cause," because it finally sees the same thing you did.

Why a recording beats a screenshot

A screenshot is a single frame with no causality. A ThinkRun recording is the sequence: the click that triggered the error, the request that failed, the console line that fired. Frontend bugs especially are visual and temporal — they don't fit in a paragraph, and a screenshot drops everything that led up to the moment.

Record once. Let the agent see exactly what happened. Get back to shipping.