CDP Support for `--user-data-dir` Parameter

[Xiechengqi]

Implementation Flow

CLI: agent-browser --cdp 9222 --user-data-dir /path/to/data
↓
flags.rs parses the argument
↓
main.rs sends via launch command: {"action":"launch","cdpPort":9222,"userDataDir":"..."}
↓
daemon.ts receives env variable, calls browser.launch()
↓
browser.ts connectViaCDP() uses args: ["--user-data-dir=..."]
↓
Playwright CDP connection includes the argument

Key Points

1. CDP mode only - Not needed for regular browser launch
2. Cross-platform path handling - Watch for path separator differences
3. Parameter passing method - Via Playwright's connectOverCDP() args option

CLI Usage Examples

  # Connect to Chrome with specific user data directory
  agent-browser --cdp 9222 --user-data-dir /app/chromium/user-data open https://x.com

  # Or separate options
  agent-browser --cdp 9222 --user-data-dir /app/chromium/user-data snapshot

[grimmjoww]

Feature: AGENT_BROWSER_CDP env var for daemon auto-launch

Currently the --cdp flag works per-command, but the daemon auto-launch (when the daemon starts and auto-opens a browser) has no way to use CDP. This means AI agents using agent-browser always get a fresh Playwright Chromium instance instead of connecting to an existing Chrome with real cookies/logins.

The Problem

When an AI agent calls agent-browser snapshot (without explicit --cdp), the daemon auto-launches a fresh Chromium browser. The agent has to log in every time, and sites with anti-bot detection (X/Twitter) block Playwright's headless Chromium.

Proposed Fix

Add AGENT_BROWSER_CDP env var support to the daemon auto-launch in daemon.ts:

// In the auto-launch block (around line 315):
const cdpEnv = process.env.AGENT_BROWSER_CDP;
const cdpUrl = cdpEnv && (cdpEnv.startsWith('ws://') || cdpEnv.startsWith('wss://') || cdpEnv.startsWith('http://') || cdpEnv.startsWith('https://')) ? cdpEnv : undefined;
const cdpPort = cdpEnv && !cdpUrl && /^\d+$/.test(cdpEnv) ? parseInt(cdpEnv, 10) : undefined;

await manager.launch({
  // ...existing options...
  ...(cdpUrl && { cdpUrl }),
  ...(cdpPort && { cdpPort }),
});

Usage

# Set env var - daemon auto-connects to running Chrome
export AGENT_BROWSER_CDP=9222

# Now any command auto-connects via CDP (no --cdp flag needed)
agent-browser snapshot -i
agent-browser open https://x.com

This is consistent with other AGENT_BROWSER_* env vars (AGENT_BROWSER_PROFILE, AGENT_BROWSER_HEADED, etc.) and makes CDP the default connection mode when set.

Tested on Windows with Chrome running --remote-debugging-port=9222 — works perfectly, agent gets real cookies and bypasses anti-bot detection.

[giulio-leone]

This is already supported via the --profile flag which maps to Playwright's launchPersistentContext(profilePath):

# Use a custom Chrome user data directory
agent-browser --profile /path/to/user-data-dir open https://example.com

# Or via environment variable
export AGENT_BROWSER_PROFILE=/path/to/user-data-dir
agent-browser open https://example.com

The --profile flag uses Playwright's launchPersistentContext which stores cookies, localStorage, cache, and all browser state in the specified directory — functionally identical to Chrome's --user-data-dir.

See PR #474 which also adds .agent-browserrc.json config file support where profile can be set as a default.

This issue can likely be closed.