Install
npx skills add eze-is/web-accessWho it is for
If you find the Agent's built-in WebSearch and WebFetch lack a dispatch policy and have no browser automation, this Agent Skill is a better fit. What it adds is a web-access policy, CDP browser actions, and accumulated site experience. The page says it works with every Agent that supports SKILL.md, and it names several products plus "etc.", so do not treat the examples as a closed roster.
People who only want text search and do not plan to turn on remote debugging can use it as a policy layer first. Once you go to CDP, you must have: Node.js 22+, and Chrome or Edge already open with remote debugging. If you only want a browser extension and do not use a SKILL.md Agent, this is not for you.
First try in 3 minutes
The official install command is written on the page as this one:
npx skills add eze-is/web-access
It writes the skill into the current Agent that supports SKILL.md. Confirm you have npx locally and that this is the environment you want to change, then run it. Do not put another installer beside it.
After install, do read-only web access first; do not control the browser immediately. Have the Agent search or fetch a public source you name, and have it say whether this time it chose WebSearch, WebFetch, curl, Jina, or a mix. Success signal: it can state the reason and return page content, instead of saying the skill is not installed. Stop here.
Only when the read-only channel is not enough and you clearly want to operate an already-open page, go into CDP. First open locally:
chrome://inspect/#remote-debugging
or
edge://inspect/#remote-debugging
Confirm remote debugging is on, then allow the Agent to act through CDP. This controls your real browser; before you run it, confirm the window and site are exactly the ones you intend.
Usage examples
Example 1: Let the skill pick web tools by situation. The table writes the capabilities as: WebSearch / WebFetch / curl / Jina / CDP, judged by situation, any mix allowed. The task should be "look this up, and tell me which channels you used". The available page text does not write the selection rules, and it does not say whether Jina needs an extra key. So what you can confirm is: these five names are in the dispatch; which one is chosen when is decided at runtime. If the Agent opens CDP and clicks the page right away, it has already gone past this read-only task.
Example 2: Use CDP on a browser that is already being debugged. Prerequisites are Node.js 22+, and the Chrome / Edge remote-debugging page already open. The purpose is browser automation, not another ordinary search. The available page text does not write clickable selectors, screenshot commands, or how to end the debug session. So the first time, ask only for one reversible read-only action (for example, look at the current page), and confirm the debug target in advance. Accumulated site experience is written as a capability, but it does not say where it is stored or how to clear it; do not let the Agent treat an unexplained memory mechanism as a database that is already configured.
Limits
CDP mode is bound to Node.js 22+ and Chrome/Edge remote debugging. If the version is too low or the inspect page is not open, do not assume browser automation is available. Ordinary WebSearch / WebFetch does not automatically mean CDP is already there.
The install path is npx skills add eze-is/web-access. Do not write a second install command. The client examples are not an authorization roster.
The first stop is a read-only search or fetch. Any CDP click, form submit, login, or write of site experience must first have the Agent repeat the window and steps it will use; you confirm, then it continues.