PlugWiki

web-access Skill (eze-is): Logged-In Browser for Agents

Gives your agent a plan for web tasks and a way to drive your everyday browser, logins included, for pages that plain fetching cannot reach.

Installs
17.2K
on skills.sh
Stars
9.3K
on GitHub
Updated
Oct 2026
last commit

web-access tells the agent how to choose between WebSearch, WebFetch, curl, the Jina page-to-Markdown service and a real browser. For the browser route it starts a small local proxy that connects to your everyday Chrome, Edge, Brave, Arc or other Chromium browser, so the agent sees pages with your logins. It opens its own background tabs, reads or clicks through them, and closes them when done.

It also keeps notes about sites in ~/.web-access/site-patterns/, can search your browser bookmarks and history, and can split research across helper agents. Before browser work it must show you this warning: some sites detect automation and may ban accounts. You turn on remote debugging in your browser first, and the SKILL.md and README are in Chinese.

Good for

  • Users who need an agent to read pages behind a login or heavy anti-scraping
  • Researchers who want first-hand sources checked, not just search snippets
  • Developers comfortable with Node.js 22+ and switching on browser remote debugging

Install web-access

Claude Code

  1. 1

    Copy this prompt.

    Install the web-access skill in Claude Code. Run this command:

npx skills add eze-is/web-access

Then check that it loaded and tell me how to use it.
  2. 2

    Paste it into Claude Code. It shows the change and asks before running it.

Requires Node.js.

Needs Node.js 22+ and a Chromium browser with remote debugging switched on (chrome://inspect/#remote-debugging). The agent then acts with that browser's logins and must show a ban-risk warning first; the README advises a spare account for social platforms. This is not a Claude app (ZIP upload) skill: its scripts need a local Node.js and your local browser.

From the vendor's docs. Source.

What's in the skill

When it loads

所有联网操作必须通过此 skill 处理，包括：搜索、网页抓取、登录后操作、网络交互等。 触发场景：用户要求搜索信息、查看网页内容、访问需要登录的网站、操作网页界面、抓取社交媒体内容（小红书、微博、推特等）、读取动态渲染页面、以及任何需要真实浏览器环境的网络任务。

Sections in SKILL.md

  1. 前置检查
  2. 浏览哲学与工作流
  3. 联网工具选择
  4. 页面就绪与完成判断
  5. 补充：本地浏览器资源
  6. 程序化操作与 GUI 交互
  7. 浏览器 CDP 模式
  8. 启动
  9. Proxy API
  10. 页面内导航
  11. 媒体资源提取
  12. 技术事实
  13. 视频内容获取
  14. 登录判断
  15. 任务结束
  16. 并行调研：子 Agent 分治策略

Key facts

Author
一泽Eze (SKILL.md metadata)
Repo
eze-is/web-access
Installs on skills.sh
17,160 (Nov 2, 2026)
GitHub stars
9,315 (repo, Nov 2, 2026)
Last commit
2026-10-14
Files
30 files in the repo, which is the skill: SKILL.md, 10 scripts, 8 script tests, 3 reference guides, a config template and README and agent notes
License (SKILL.md)
MIT
Version (SKILL.md)
2.6.0
Needs
Node.js 22+; Chromium browser with remote debugging on; sqlite3 for history search
Works in
Claude Code, Cursor, Gemini CLI, Codex CLI (any SKILL.md agent)
Language
SKILL.md and README are in Chinese

What it includes

  • Pre-check: reuses a running proxy or starts a background one, needs Node.js 22+

    Exit 0 means ready, exit 2 means first-time browser setup is needed, exit 1 prints the error. It starts the proxy as a detached background process.

    Runs code · check-deps.mjs

  • Picks WebSearch, WebFetch, curl, Jina or the browser by task

    Search to find sources, fetch to extract from known URLs, curl for raw HTML, browser for logins and anti-scraping sites. Jina (r.jina.ai, 20 requests a minute) is an optional page-to-Markdown step.

    Network · Tool selection

  • Local HTTP API: /new, /eval, /click, /clickAt, /clickXY, /scroll, /screenshot, /close

    Runs on port 3456 by default, bound to 127.0.0.1. The agent calls it with curl. /eval runs JavaScript in a page.

    Runs code · CDP proxy API

  • /frames lists iframes, /eval?frame= runs inside one, /setFiles sets a file input

    /setFiles points a page's file input at a local file path.

    Files · Frames and uploads

  • Works in background tabs it creates, leaves your tabs alone, closes its own after

    Tabs it opens that sit idle for 15 minutes are cleaned up by the proxy.

    Open only its own tabs

  • Searches Chrome and Edge bookmarks and history by keyword, time or visit count

    Copies the browser History database to a temp file and queries it with the sqlite3 command. Used when you refer to pages you visited or internal company systems.

    Runs code · find-url.mjs

  • status, stop and restart for the shared proxy

    Refuses to stop while requests or agent tabs are active.

    Runs code · proxy-maintenance.mjs

  • Reads and writes per-domain notes in ~/.web-access/site-patterns/

    After a successful browser task it may record URL patterns and traps. Your browser choice is kept in ~/.web-access/config.env.

    Write · Site experience

  • Splits independent research goals across helper agents sharing one proxy

    Each helper opens and closes its own tabs.

    Network · Parallel research

  • Goes to first-hand sources and says so when only second-hand reports exist

    Search results are for finding sources, not proving a claim.

    Network · Fact checking

From the official docs

Permissions

High risk

The skill lets an agent drive your everyday browser with your logins through a background proxy, and sites may ban accounts that automation touches. Use a separate profile or spare accounts and stop the proxy when you finish.

What it can do

  • Read your data
  • Create and change things
  • Delete things
  • Read-only mode

Watch out for

  • The agent acts with your browser logins

    The proxy connects to your everyday browser and can run JavaScript, click, upload files and take screenshots in tabs it opens. Our advice: use a separate browser profile or spare accounts. github.com

  • Sites may ban accounts for automation

    The skill makes the agent show a ban-risk warning, and the README strongly advises a spare account for social platforms. github.com

  • A background proxy keeps running

    check-deps.mjs starts the proxy as a detached process and SKILL.md says to leave it running. Stop it with proxy-maintenance.mjs when you finish. github.com

  • It can read your bookmarks and history

    find-url.mjs searches Chrome and Edge bookmarks and a temporary copy of the History database. Tell the agent when you do not want that. github.com

  • Optional Jina step sends URLs to a third party

    SKILL.md describes r.jina.ai as a third-party service that turns pages into Markdown, so the page addresses go to it. github.com

  • It writes notes under ~/.web-access

    Browser choice and per-site notes are saved there, outside the skill folder. github.com

  • Plugin route adds the whole repo as one skill

    The README's Claude Code plugin install uses .claude-plugin/plugin.json and marketplace.json, which list the repo root as the skill (skills: ["./"]). The two manifests declare no hooks or MCP servers. github.com

Technical details
Talks toAny site the agent opens. r.jina.ai (optional Jina step). local proxy on 127.0.0.1:3456
Scopes

  • Remote debugging of your everyday browser

Questions

What does the web-access skill do?

It makes the agent pick the right way to get web information and, when needed, drive your real browser through a local proxy. That reaches pages behind logins, dynamic pages, and sites like Xiaohongshu that block simple fetching.

How do I install it?

Use the install box above. The README recommends npx skills add eze-is/web-access and also lists a Claude Code plugin route and a git clone into ~/.claude/skills/web-access. Then turn on "Allow remote debugging" at chrome://inspect/#remote-debugging in your browser. The box has no Claude app upload tab because the scripts need a local Node.js and your local browser.

What do I need?

Node.js 22+ (it uses the built-in WebSocket) and a Chromium-based browser with remote debugging switched on. find-url.mjs also needs the sqlite3 command. No API key is named in SKILL.md.

Does it work in Claude Code, Cursor, Gemini CLI and Codex?

The README says it works with any agent that supports SKILL.md and names Claude Code, Cursor, Gemini CLI and Codex CLI.

Can it use my logged-in accounts?

Yes, that is the point. It shares your everyday browser, so the agent acts with your sessions. It is told to work in its own tabs, but anything it opens can use your logins.

Can I get banned?

It is possible. The SKILL.md makes the agent warn you that some sites detect browser automation and may ban accounts. The README strongly advises using a spare account for social platforms.

What does it do with my bookmarks and history?

Only when you mention pages you visited or internal systems, it can search bookmarks and history in Chrome and Edge with find-url.mjs. It copies the History database to a temporary file to read it.

Is it safe to run?

It carries real risk. It runs a background proxy that can run JavaScript in pages, click, upload files and take screenshots, all with your browser logins. Our advice: use a separate browser profile or spare accounts, review what the agent plans to do, and stop the proxy when you are done.

Information

updated Nov 2, 2026

Repository
eze-is/web-access

eze-is/web-accessskills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 8
  1. web-access README · github.com
  2. plugin.json · github.com
  3. web-access SKILL.md · github.com
  4. skills.sh listing · skills.sh
  5. eze-is/web-access · github.com
  6. cdp-proxy.mjs · github.com
  7. find-url.mjs · github.com
  8. check-deps.mjs · github.com

Updated Nov 2, 2026.