Drive My Browser, Take Two
Let Your AI Drive Chrome, Signed In as You (Take Two)
Back in June I wrote Drive My Browser: a tiny skill that let my AI agent use websites as me, with my logins intact.
It worked great. Then a Chrome update came along, and my agent got a sign-in screen on every site. The trick was to give the agent a copy of my Chrome profile, and a copy now comes up signed out of everything. I don't know why.
So the new version doesn't copy anything. The agent gets its own Chrome profile, and I sign in there once.
Step 1: Install agent-browser
npm i -g agent-browser
You want version 0.42 or newer.
Step 2: Add the script
Save this as ~/.claude/skills/drive-my-chrome/agent-chrome and make it
executable with chmod +x. It's written for macOS.
Show agent-chrome
#!/bin/bash
#
# agent-chrome: a signed-in Chrome for agents to drive.
#
# Runs real Google Chrome on its own profile folder, which is kept between
# runs. You sign in there once per site with `agent-chrome login`. Agents then
# attach with `agent-browser --cdp <port>`.
#
# Chrome is started here, not by agent-browser, so that it uses the real cookie
# key from the system keychain and can read the saved sign-ins.
#
# agent-chrome start start headless Chrome if needed; print the port
# agent-chrome status say whether it is running, and how
# agent-chrome stop stop it
# agent-chrome login [url...] stop it, then open a visible window to sign in
#
# AGENT_CHROME_DIR and AGENT_CHROME_PORT override the folder and the port.
set -u
DIR="${AGENT_CHROME_DIR:-$HOME/.agent-browser/agent-chrome}"
PORT="${AGENT_CHROME_PORT:-9333}"
APP="Google Chrome"
BIN="/Applications/Google Chrome.app/Contents/MacOS/Google Chrome"
# The main Chrome process running on this folder, if any.
chrome_pid() {
ps -axo pid=,command= | grep -F -- "--user-data-dir=$DIR" | grep -F "$BIN" | grep -v -F "Helper" | awk 'NR==1 {print $1}'
}
chrome_mode() {
ps -o command= -p "$1" 2>/dev/null | grep -q -- "--remote-debugging-port=$PORT" && echo "headless" || echo "visible"
}
port_answers() {
curl -s -o /dev/null --max-time 2 "http://127.0.0.1:$PORT/json/version"
}
stop_chrome() {
local pid
pid="$(chrome_pid)"
[ -z "$pid" ] && return 0
kill "$pid" 2>/dev/null
for _ in $(seq 1 20); do
[ -z "$(chrome_pid)" ] && return 0
sleep 0.5
done
echo "agent-chrome: Chrome (pid $pid) did not stop" >&2
return 1
}
case "${1:-}" in
start)
pid="$(chrome_pid)"
if [ -n "$pid" ] && [ "$(chrome_mode "$pid")" = "visible" ]; then
echo "agent-chrome: the sign-in window is still open. Quit it, then run start again." >&2
exit 1
fi
if [ -z "$pid" ]; then
if [ ! -d "$DIR" ]; then
echo "agent-chrome: no profile yet. Run: agent-chrome login" >&2
exit 1
fi
if port_answers; then
echo "agent-chrome: port $PORT is in use by something else. Set AGENT_CHROME_PORT." >&2
exit 1
fi
# `open -n` hands Chrome to the system, so it outlives this shell.
open -na "$APP" --args --user-data-dir="$DIR" --remote-debugging-port="$PORT" --headless=new \
--no-first-run --no-default-browser-check --window-size=1440,1600 about:blank
fi
for _ in $(seq 1 30); do
if port_answers; then
echo "$PORT"
exit 0
fi
sleep 0.5
done
echo "agent-chrome: Chrome did not open port $PORT" >&2
exit 1
;;
status)
pid="$(chrome_pid)"
if [ -z "$pid" ]; then
echo "stopped"
else
echo "running ($(chrome_mode "$pid")), pid $pid, port $PORT"
fi
;;
stop)
stop_chrome
;;
login)
shift
stop_chrome || exit 1
mkdir -p "$DIR" && chmod 700 "$DIR"
# A plain window with no debugging port, so sign-in pages see an ordinary browser.
open -na "$APP" --args --user-data-dir="$DIR" --no-first-run --no-default-browser-check "$@"
echo "A Chrome window is open. Sign in, choose \"Use Chrome Without an Account\" if Chrome asks, then quit that window."
;;
*)
echo "usage: agent-chrome start|status|stop|login [url...]" >&2
exit 2
;;
esac
Don't skip the script. When agent-browser starts Chrome by itself, Chrome
deletes the sign-ins you just saved. I don't fully know why that is either.
The script starts Chrome the ordinary way, and agent-browser attaches to it.
Step 3: Add the skill
Save this as ~/.claude/skills/drive-my-chrome/SKILL.md:
Show SKILL.md
---
name: drive-my-chrome
description: Use when browser automation must be signed in as the user — phrases like "use my chrome", "logged in as me", "browse as myself", or any site behind a login or SSO the user already holds. Drives real Google Chrome, headless, on a dedicated profile the user has signed in to. Reach for it instead of a plain `agent-browser open` whenever the task implies the user's own session.
---
# drive-my-chrome
Drive **real Google Chrome**, already signed in as the user, with agent-browser.
Headless, no prompts, and safe for several agents at once.
## How it works
- Agents get their own Chrome profile folder, `~/.agent-browser/agent-chrome`.
The user signs in there once per site. It is separate from the user's
everyday Chrome, which is never touched.
- The `agent-chrome` script next to this file starts real Chrome on that folder
with a debugging port. agent-browser attaches to it with `--cdp`.
- Chrome is always started by `agent-chrome`, never by agent-browser. Chrome
encrypts cookies with a key from the system keychain, and agent-browser
launches Chrome with a stand-in key, which makes Chrome discard the saved
sign-ins.
## The recipe
### 1. Start Chrome and open your own tab
```bash
PORT=$(~/.claude/skills/drive-my-chrome/agent-chrome start) || exit 1
AGENT_BROWSER_SESSION=my-task agent-browser --cdp "$PORT" --pin-tab tab new "https://example.com/"
AGENT_BROWSER_SESSION=my-task agent-browser --cdp "$PORT" --pin-tab get title
```
- `start` is safe to repeat. It starts Chrome if needed and prints the port.
- Put the session name, `--cdp "$PORT"` and `--pin-tab` on **every** command.
`--pin-tab` keeps the session on the tab it opened.
- Begin with `tab new`. Each agent uses its own session name and its own tab,
so several can share this Chrome at the same time. `tab list` shows every
agent's tabs; touch only your own.
- For screenshots, set the size once after `tab new`: `set viewport 1440 1600`.
### 2. Check that you are signed in
Look at the page title, the address, or the account name in a snapshot before
relying on the session. A sign-in page means the profile has no session for
that site yet.
### 3. On a sign-in page, stop and ask
Never type credentials, and never go looking for a password. Open a visible
window on the same profile and ask the user to sign in:
```bash
~/.claude/skills/drive-my-chrome/agent-chrome login "https://example.com/"
```
Ask the user to:
1. Sign in on that page.
2. Choose **Use Chrome Without an Account** if Chrome offers to sign the
browser itself in.
3. Quit that window when done.
Then run `agent-chrome start` again and reopen your tab.
`login` stops the headless Chrome first, which disconnects any other agent
using it, so run it only when nothing else is in progress.
### 4. Clean up
```bash
AGENT_BROWSER_SESSION=my-task agent-browser --cdp "$PORT" --pin-tab tab close
AGENT_BROWSER_SESSION=my-task agent-browser close
```
- Close the tab you opened, then disconnect. Chrome keeps running for other
agents.
- Use plain `close`. `agent-browser close --all` ends every agent-browser
session on the machine.
- `agent-chrome stop` shuts Chrome down when nothing else needs it.
## The script
| Command | What it does |
|---|---|
| `agent-chrome start` | Starts headless Chrome if it is not running, and prints the port |
| `agent-chrome status` | Says whether Chrome is running, and whether headless or visible |
| `agent-chrome stop` | Stops it |
| `agent-chrome login [url...]` | Stops it, then opens a visible window for signing in |
`AGENT_CHROME_DIR` changes the profile folder and `AGENT_CHROME_PORT` the port
(9333 by default).
## Good to know
- **Sign-ins are per site and they expire.** A site the user is signed in to in
their everyday Chrome is not signed in here until they do step 3 for it, and
each site ends its sessions on its own schedule.
- **One Chrome per profile folder.** While the visible `login` window is open,
`start` refuses to run, and the other way round.
- **Sign-in is always done by the user, in the `login` window.** Identity
providers often refuse a browser that is being automated.
- **Start with the flags.** If a session was first used without `--cdp`, close
that session and begin again with the flags on the first command.
Step 4: Sign in once
~/.claude/skills/drive-my-chrome/agent-chrome login "https://example.com/"
A fresh Chrome window opens. Sign in to the site, then quit the window. If Chrome asks to sign the browser itself in, pick Use Chrome Without an Account.
When a site signs you out later, run this again.
Step 5: Just ask
"drive my chrome, open my dashboard and tell me what changed since yesterday"
The agent runs Chrome in the background, already signed in. You won't see a window.
That's it. 🤖