Drive My Browser

Let Your AI Drive Your Chrome — Logged In as You

Want your AI agent to actually use a website for you — search, click, fill forms — but as you, with your accounts already logged in? You need basically two things: agent-browser installed, and a little skill file that teaches your agent the trick. That's it.

Step 1: Install agent-browser

npm i -g agent-browser && agent-browser install

This is the only real install. agent-browser is a CLI that lets an AI control a browser.

Step 2: Add the skill

agent-browser can drive a browser, but on its own it opens a blank test browser where you're logged into nothing. Two small gotchas get in the way:

• Modern Chrome won't let a robot remote-control your everyday profile.
• The bundled test browser can't read your saved logins (they're locked to real Chrome).

The fix is one tidy command — and instead of memorizing it, you hand it to your agent as a skill. Save this as ~/.claude/skills/drive-my-chrome/SKILL.md:

---
name: drive-my-chrome
description: Drive the user's real, logged-in Chrome with one of their existing
  Chrome profiles via agent-browser. Use when the user wants browser automation
  that reuses their own login/cookies — "use my chrome profile", "my default
  profile", "logged in as me", "browse as myself".
allowed-tools: Bash(agent-browser:*), AskUserQuestion
---

# drive-my-chrome

Launch the user's **real** Google Chrome with one of their existing profiles,
so logins/cookies are intact, and drive it with agent-browser.

## Steps

1. List profiles, so you know what's available:
   ```bash
   agent-browser profiles
   ```
   Output maps folder → account, e.g. `Profile 1  (you@work.com)`.

2. Pick the profile. If the user named one clearly, use it. Otherwise **ask**
   with AskUserQuestion — "my default profile" is ambiguous (the `Default`
   folder, or the account they live in?). Pass `--profile` the **folder name**
   (`Default`, `Profile 1`, …), not the email.

3. Launch real Chrome, visible, in a named session:
   ```bash
   REALCHROME="/Applications/Google Chrome.app/Contents/MacOS/Google Chrome"
   AGENT_BROWSER_SESSION=mychrome agent-browser \
     --executable-path "$REALCHROME" \
     --profile "Profile 1" \
     --headed \
     open "https://www.google.com"
   ```
   (Linux: `/usr/bin/google-chrome`. Windows:
   `C:\Program Files\Google\Chrome\Application\chrome.exe`.)

4. Drive it — reuse the same session name for every command:
   ```bash
   export AGENT_BROWSER_SESSION=mychrome
   agent-browser snapshot -i        # see the page; elements show as @e1, @e2…
   agent-browser fill @e16 "wow"    # type into the search box
   agent-browser press Enter
   ```

5. Confirm you're logged in, then clean up when done:
   ```bash
   agent-browser snapshot -i | grep -i "google account"   # should show your account
   agent-browser close --all
   ```

## Notes
- `--executable-path` = real Chrome (so your cookies decrypt). `--profile`
  copies the profile to a temp dir so Chrome will accept remote control. Get
  both right and you're logged in.
- Each launch is a throwaway copy — new logins/bookmarks don't sync back to
  your real Chrome.
- Google specifically may still bounce automation to a CAPTCHA. Most other
  sites stay logged in fine.

Step 3: Just ask

That's the whole setup. Now you talk to your agent in plain English:

"drive my chrome, go to google and search for wow"

The skill kicks in, asks which profile you mean if it's unsure, opens your real Chrome logged in as you, and does the thing. 🤖

That's really it — install agent-browser, drop in the skill, and ask. Everything else is the agent's job.

Back to Index