Chrome Cdp

Chrome CDP

This built-in skill uses MCP Gateway's bundled cdp.mjs runner. It talks directly to Chrome DevTools Protocol over WebSocket with no third-party browser automation bridge.

Default Browser Mode

By default, the gateway starts or reuses a visible Chromium-family browser in managed launch mode.

• It does not attach to the user's already-open Chrome unless the gateway process is started with CDP_BROWSER_MODE=attach.
• It uses a gateway-managed runtime directory for target caches, per-tab daemons, and CDP state.
• It sets CDP_PROFILE_MODE=persistent by default.
• It sets CDP_USER_DATA_DIR to the gateway-owned persistent profile path: mcp-gateway/builtin-skills/chrome-cdp/chrome-user-data under the OS local data directory.
• Cookies, cache, localStorage, and login sessions are reused across commands, browser restarts, and gateway restarts unless the site expires them or the profile directory is deleted.

Cache and Login State

Default behavior is to reuse the existing browser cache/profile. Do not create a fresh profile just because a new task starts, a new page is opened, or the browser was closed and reopened.

Use the default persistent profile when the user asks to:

• open a page
• continue browsing
• debug a site
• use an already logged-in session
• reopen the browser later

Only use a new cache/profile when the user explicitly asks for a clean browser, fresh cache, separate login, separate account, temporary session, or similar. In that case, start the gateway with a different CDP_USER_DATA_DIR, or set CDP_PROFILE_MODE=empty for a blank managed profile. Do not delete or overwrite the default persistent profile unless the user explicitly asks to clear saved browser data.

Commands

Short form accepted by the gateway:

open https://example.com
list
snap
eval "document.title"
netclear
net api/chat
netget 1234.56 --request-file "D:\tmp\chat-request.txt" --response-file "D:\tmp\chat-response.txt"
stop

You may also pass the script form; the gateway still runs its bundled script:

node cdp.mjs open https://example.com

Targets

Page-scoped commands can take a target prefix from list.

list
snap <target>
html <target> ".message-list"
click <target> "button[type=submit]"

For convenience, open, launch, and list maintain an active target. These commands may omit the target and use the active page:

snap
eval "document.title"
netclear
net chat
netget <request-id>

If multiple pages are open and no active target is known, run list and pass the target prefix explicitly.

Navigation and DOM

launch [url]                    # start or reuse managed browser
open [url]                      # open a tab and select it as active
list                            # list open pages and target prefixes
snap [target]                   # accessibility tree snapshot
html [target] [selector]        # full page or element HTML
nav <target> <url>              # navigate target to URL
eval [target] <js>              # evaluate JavaScript in the page
evalraw <target> <method> [json]# raw CDP command passthrough

eval accepts expressions or function/IIFE source. When calling this tool through MCP Gateway on Windows, quote the whole JavaScript expression as one argument.

Interaction

click <target> <selector>       # CSS selector
clickxy <target> <x> <y>        # CSS pixel coordinates
type <target> <text>            # Input.insertText at current focus
loadall <target> <selector> [ms]
shot [target] [file] [--inline-base64]

Use type after focusing an input with click or clickxy. It uses CDP input events and works better than DOM assignment for cross-origin iframes.

Screenshot output modes

• Default writes the PNG to disk under the runtime directory (or [file]) and prints its absolute path plus DPR / coordinate mapping. Choose this when the image is large, must be reused across turns, or should not load the model's context window.
• shot [target] [file] --inline-base64 skips the file write and returns a mimeType: image/png base64 payload inside stdout (as JSON after the [inlineBase64 json] marker), plus the same DPR / coordinate mapping lines. Choose this when a multimodal model should consume the image directly in the current turn.