| name | operating-cdp-cli |
| description | Operates the cdp CLI for browser navigation, element interaction, JavaScript evaluation, browser attachment, and interactive REPL workflows. Use when the task involves driving Chrome or Chromium with the repo's cdp command, exploring available commands, or choosing between interactive and scripted execution. |
Operating CDP CLI
Use this skill when the task is primarily about running or explaining the cdp command.
Quick start
- Build with
go build -o cdp ./cmd/cdp
- Use interactive mode for exploration and ad hoc debugging
- Use
cdp run when the task should be reproducible as a script
Use this skill for
- connecting to a local or remote browser
- navigating pages and interacting with DOM elements
- evaluating JavaScript or extracting page state
- explaining common
cdp commands and aliases
Live agent loop
For live browser work, prefer cdp --mcp and use a screenshot-first loop:
- Observe with
screenshot, usually with annotation enabled, and use page_snapshot when you need accessible element refs.
- Act with
click or type_text using @ref values first, CSS selectors when stable, and coord:x,y viewport clicks when the target is visible but not represented well in the accessibility tree.
- Verify with another
screenshot after every meaningful action; use get_page_content, get_console, and get_errors for text and failure context.
- Refresh refs with
page_snapshot after navigation or large DOM changes; stale refs should not be debugged manually first.
- Use specialized tools before raw protocol calls:
list_frames/switch_frame, handle_dialog, upload_file, storage/cookie tools, network log tools, screenshots, and PDFs.
- Use
raw_cdp only as an escape hatch for a missing tool or protocol-specific diagnosis.
Treat screenshots as the source of truth for visible state. Text snapshots are complements, not replacements, because many browser failures are layout, focus, overlay, iframe, or dialog problems.
Read next
