Chrome Cdp

Chrome CDP

Lightweight Chrome DevTools Protocol CLI. Connects directly via WebSocket — no Puppeteer, works with 100+ tabs, instant connection.

Modes

Default — connect to the user's real Chrome

When invoked without flags, scripts/cdp.mjs auto-discovers the running Chrome instance from the standard profile location (~/Library/Application Support/Google/Chrome/... on macOS, ~/.config/google-chrome/... on Linux, %LOCALAPPDATA%\Google\Chrome\User Data\... on Windows). Remote debugging must be enabled in the user's Chrome (toggle once at chrome://inspect/#remote-debugging).

Use this mode when the user wants to inspect a page they already have open.

--isolated — sandboxed Chrome with a dedicated profile

When the user passes --isolated, drive a separate Chrome instance using --user-data-dir=$HOME/.chrome-debug-profile and --remote-debugging-port=9222. The user's real Chrome profile is untouched (no extensions, no logins, no cookies leak in either direction).

Two-step protocol for every --isolated invocation:

1. Run the launcher (idempotent — exits fast if the sandboxed Chrome is already up):

   scripts/chrome-debug.sh
   

2. Run any cdp.mjs command with CDP_PORT_FILE pointing at the isolated profile, e.g.:

   CDP_PORT_FILE="$HOME/.chrome-debug-profile/DevToolsActivePort" scripts/cdp.mjs list
   CDP_PORT_FILE="$HOME/.chrome-debug-profile/DevToolsActivePort" scripts/cdp.mjs shot <target>
   

The env var overrides auto-discovery (see getWsUrl() in scripts/cdp.mjs), forcing the connection to the sandboxed instance. Without the env var, the same cdp.mjs would connect to the user's real Chrome.

Both scripts are referenced relative to the skill directory — Claude Code resolves them against the skill's root.

Argument parsing

Before running any command, scan the user's request for --isolated:

• Present → follow the Isolated protocol above for every cdp.mjs call in this turn.
• Absent → call scripts/cdp.mjs <command> directly without the env var.

Never silently pick isolated mode. If the user says "screenshot tab X", that's default mode. If they say "use a clean profile", "don't touch my Chrome", or pass --isolated, that's isolated mode. When ambiguous, ask.

Prerequisites

• Node.js 22+ (uses built-in WebSocket).
• For default mode: enable remote debugging once at chrome://inspect/#remote-debugging in the user's Chrome.
• For --isolated: nothing — scripts/chrome-debug.sh will launch Chrome on demand.

Commands

All commands use scripts/cdp.mjs. The <target> is a unique targetId prefix from list; copy the full prefix shown in the list output (for example 6BE827FA). The CLI rejects ambiguous prefixes.

List open pages

scripts/cdp.mjs list

Take a screenshot

scripts/cdp.mjs shot <target> [file]    # default: screenshot-<target>.png in runtime dir

Captures the viewport only. Scroll first with eval if you need content below the fold. Output includes the page's DPR and coordinate conversion hint (see Coordinates below).

Accessibility tree snapshot

scripts/cdp.mjs snap <target>

Evaluate JavaScript

scripts/cdp.mjs eval <target> <expr>

Watch out: avoid index-based selection (querySelectorAll(...)[i]) across multiple eval calls when the DOM can change between them (e.g. after clicking Ignore, card indices shift). Collect all data in one eval or use stable selectors.

Other commands

scripts/cdp.mjs html    <target> [selector]   # full page or element HTML
scripts/cdp.mjs nav     <target> <url>         # navigate and wait for load
scripts/cdp.mjs net     <target>               # resource timing entries
scripts/cdp.mjs click   <target> <selector>    # click element by CSS selector
scripts/cdp.mjs clickxy <target> <x> <y>       # click at CSS pixel coords
scripts/cdp.mjs type    <target> <text>         # Input.insertText at current focus; works in cross-origin iframes unlike eval
scripts/cdp.mjs loadall <target> <selector> [ms]  # click "load more" until gone (default 1500ms between clicks)
scripts/cdp.mjs evalraw <target> <method> [json]  # raw CDP command passthrough
scripts/cdp.mjs open    [url]                  # open new tab (each triggers Allow prompt)
scripts/cdp.mjs stop    [target]               # stop daemon(s)

Coordinates

shot saves an image at native resolution: image pixels = CSS pixels × DPR. CDP Input events (clickxy etc.) take CSS pixels.

CSS px = screenshot image px / DPR

shot prints the DPR for the current page. Typical Retina (DPR=2): divide screenshot coords by 2.

Tips

• Prefer snap over html for page structure — the daemon always returns it compact.
• Use type (not eval) to enter text in cross-origin iframes — click/clickxy to focus first, then type.
• Chrome shows an "Allow debugging" modal once per tab on first access. A background daemon keeps the session alive so subsequent commands need no further approval. Daemons auto-exit after 20 minutes of inactivity.
• In --isolated mode the sandboxed Chrome starts with no extensions, no logins, and a blank profile. Re-launching it (scripts/chrome-debug.sh) is a no-op while it's running.
• Batching: because the daemon holds the session open, chaining several cdp.mjs calls with && in one Bash call is already a batch — click throws when its selector is missing, so a bad step stops the chain instead of mis-clicking the next one. There's no wait command; poll for a post-action element with eval's awaitPromise: true support instead of a fixed sleep.