Chrome CDP
Skill
After explicit user approval, guides an agent to inspect, debug, or interact with an existing local Chrome-family tab through a lightweight CDP CLI.
What it does
Only after explicit user approval, Chrome CDP guides an agent through inspecting, debugging, or interacting with a page open in the user's local Chrome-family browser.
Its Node command-line interface connects over the Chrome DevTools Protocol and supports tab listing, screenshots, accessibility snapshots, HTML or JavaScript inspection, navigation, and controlled input.
When to reach for it
Invocation: First ask for approval to inspect, debug, or interact with a page open in Chrome; invoke this skill only after the user explicitly approves that browser task.
- The user approves inspection of the current state of an already-open page
- List targets, select the approved tab, then capture the smallest useful snapshot or screenshot.
- The user approves debugging a live DOM, accessibility, or resource-timing problem
- Run the narrow read-only command that answers the question before considering any interaction.
- The user approves navigation, clicking, or typing in a named page
- Act only in the approved target and report the resulting page state.
What it adds
- An attach-oriented CDP workflow for an existing local browser session instead of a newly launched automation profile.
- A fail-closed Lightpanda fallback message that requires fresh user approval before using a separate browser backend.
How it fits
No related catalog surfaces are recorded.
Common questions
- Does loading the skill grant permission to inspect Chrome?
- No. Explicit user approval for the intended inspection, debugging, or interaction is required before any command runs. Chrome's own Allow debugging prompt is an additional browser gate, not a substitute for that approval.
- What happens if Lightpanda cannot perform an operation?
- The command stops and explains that another browser has separate cookies, login, storage, DOM, and workflow state. A fallback configuration changes suggestion text only; execution still waits for explicit user approval.
It’s working if
- The user explicitly approved the named browser task before any CDP command ran.
- The selected target belongs to the approved browser and page, and output answers only the approved task.
- An unsupported Lightpanda operation stops until the user separately approves a fallback browser.
Source and setup
Provenance
- Project
- Chrome CDP
- Origin
- Petr Baudis
- Current maintainer
- Vedang Manerikar
- Relationship
- fork
- Installed source
git:github.com/vedang/chrome-cdp-skill- Source path
- View source (
skills/chrome-cdp/SKILL.md) - Compatibility
- Agent Skills
- Node.js 22+
- Chrome-family browser with remote debugging enabled
- License
- MIT
- License publication scope
- Source link, independently written summary, original-author and fork attribution, and reviewed behavior claims; retain MIT attribution.
- Reviewed revision
24d2354b1041a51219d3e42013162e83b787e740- Catalog revision
e0c035a297462c519315abfa1af31a1ed3bd9eef
Setup notes
- Created by Petr Baudis; this installed MIT-licensed fork is maintained by Vedang Manerikar.
- Requires Node runtime version 22 or newer and a supported browser with remote debugging enabled.
- Treat authenticated tabs as sensitive: describe the intended access and obtain explicit user approval before listing, reading, or changing browser state.
- Behavior claims were reviewed against the installed commit rather than a moving branch.