Chrome for Agents

Give your AI agent full access to your live Chrome browser — existing tabs, saved logins, cookies, and all — without restarting Chrome, creating separate profiles, or re-authenticating anything.

Most AI agent browser tools launch a headless or isolated browser instance. That means:

• No saved passwords or cookies — you re-login to everything
• No access to already-open tabs
• Bot detection flags headless Chrome immediately (eBay, Amazon, etc.)
• You can't see what the agent is doing or intervene

This setup uses the Chrome DevTools MCP server to attach your agent to your already-running Chrome session. You keep full visibility and control — the agent acts inside your real browser.

Because the browser is yours — visible, real, and running — you can jump in at any moment. This is the biggest advantage over headless approaches:

• CAPTCHAs — Agent hits a CAPTCHA? You click the checkbox. It's sitting right there in your browser.
• Anti-bot challenges — Cloudflare, Incapsula, and similar protections see a real Chrome session with real cookies and browsing history. No fingerprint spoofing needed.
• Credential entry — 2FA prompts, password managers, biometric auth — your meat-sack appendages handle what silicon can't.
• Approval gates — See exactly what the agent is about to submit before it happens. Your mouse is right there.
• Recovery — Agent gets stuck on an unexpected modal or popup? You dismiss it and let it continue.

This isn't a workaround — it's the architecture. The agent drives, you co-pilot. Headless browsers try to solve these problems with increasingly fragile automation. We just let the human be human.

Agent (Hermes/Claude/etc.)
    ↕ MCP Protocol
chrome-devtools-mcp server (Node.js)
    ↕ Chrome DevTools Protocol
Your running Chrome (v144+)
    ↕
Your real tabs, cookies, logins

The chrome-devtools-mcp server is an official Google project (ChromeDevTools/chrome-devtools-mcp) that bridges the MCP (Model Context Protocol) to Chrome's DevTools Protocol. The --autoConnect flag (Chrome M144+) lets it attach to your running Chrome without requiring a debug port launch.

• Chrome 144+ — check at chrome://version
• Node.js 20.19+ — node --version
• An MCP-compatible agent (Hermes, Claude Desktop, Cursor, etc.)

Run this one-liner:

curl -fsSL https://raw.githubusercontent.com/telos-oc/chrome-for-agents/main/install-hermes.sh | bash

What it does:

• Checks Python 3, Node.js 20.19+, npx, Hermes, and Chrome/Chromium 144+
• Adds or updates the chrome-devtools MCP server in ~/.hermes/config.yaml
• Verifies the config was written correctly
• Verifies chrome-devtools-mcp@latest resolves via npx
• Restarts the Hermes gateway

The installer is idempotent — re-running it updates the managed chrome-devtools entry without duplicating it.

If you prefer to inspect the script before running it:

curl -fsSL https://raw.githubusercontent.com/telos-oc/chrome-for-agents/main/install-hermes.sh -o install-hermes.sh
bash install-hermes.sh

1. Open Chrome
2. Navigate to chrome://inspect/#remote-debugging
3. Enable the "Discover network targets" toggle (if not already on)

That's it. No Chrome restart needed, no special launch flags, no separate profile.

Add to ~/.hermes/config.yaml under mcp_servers:

mcp_servers:
  chrome-devtools:
    command: npx
    args:
      - -y
      - chrome-devtools-mcp@latest
      - --autoConnect
      - --no-usage-statistics
    timeout: 120
    connect_timeout: 60

Then restart the gateway:

hermes gateway restart

1. Open Chrome to chrome://inspect/#remote-debugging
2. Make sure Discover network targets is enabled
3. Keep at least one Chrome tab open
4. Ask Hermes to list browser tabs
5. When Chrome shows the consent prompt, click Allow

If Hermes returns your existing Chrome tabs, you're connected.

On first use, Chrome will show a consent prompt asking you to approve the remote debugging attachment. Click Allow.

Note: This approval prompt may reappear after restarting your Mac or Chrome — it's a one-time consent per session, not a permanent grant. Just click Allow again when it pops up.

This is your security gate — the agent can't connect to your browser without your explicit approval each session.

Once connected, the agent gets ~29 tools including:

# Had to fully quit Chrome first
/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome \
  --remote-debugging-port=9222 \
  --user-data-dir=/tmp/chrome-debug-profile &

Problems:

• Chrome must be fully closed first — can't run alongside your normal browser
• --user-data-dir is mandatory (Chrome's multi-profile chooser locks the default directory)
• Separate profile = no existing logins, no cookies, no saved passwords
• Had to use 127.0.0.1 not localhost (macOS IPv6 resolution issue)
• Manual credential management via .env files for sites with expiring cookies

mcp_servers:
  chrome-devtools:
    command: npx
    args: ["-y", "chrome-devtools-mcp@latest", "--autoConnect"]

What changed:

• No Chrome restart — attaches to your running session
• All your existing tabs, cookies, and logins are immediately available
• User approval prompt provides a security gate
• Official Google project, well-maintained
• Works with Chrome, Brave, Edge, and other Chromium browsers

• Chrome 144+ required — older versions don't support --autoConnect
• Left-click only — no right-click or middle-click
• Text entered all at once — no "type slowly" option
• No batch actions — each interaction is a separate tool call
• No download/PDF interception — browser handles these natively
• Chrome must stay running — if Chrome closes, all browser tools fail
• Snapshot refs change — element UIDs update after navigation or major DOM changes; re-snapshot after actions
• User must approve — first attach requires manual consent in Chrome

This setup gives the agent full access to your real browser session. That means:

• The agent can see your open tabs and their content
• The agent acts with your cookies and authentication state
• Destructive actions (closing tabs, navigating away, submitting forms) affect your real browser
• You should monitor the browser during sensitive operations

The Chrome consent prompt on each session is your primary security gate. If you don't approve, the agent can't connect.

1. Pin important tabs — the agent can identify tabs by title/URL, so descriptive tab names help
2. Keep Chrome visible — part of the value is watching what the agent does
3. Use for bot-sensitive sites — this approach passes anti-bot checks that headless browsers fail (eBay, Amazon, etc.)
4. The --no-usage-statistics flag — prevents the MCP server from sending telemetry to Google

Agent can't find Chrome:

• Verify remote debugging is enabled at chrome://inspect/#remote-debugging
• Check Chrome version is 144+: chrome://version

Connection hangs:

• Look for the Chrome consent prompt — it might be hidden behind other windows
• The consent prompt reappears after Mac/Chrome restarts — click Allow again
• Try restarting the MCP server (restart your agent's gateway)

Agent reports "no pages":

• Make sure Chrome has at least one tab open
• The MCP server needs a page context to attach to

• chrome-devtools-mcp GitHub — the MCP server
• Model Context Protocol — the protocol spec
• Hermes Agent — the agent framework this was built for