Skip to content

Browser MCP

Some posting pages show their text only after a browser has run their scripts. A plain page fetch (see Web search) then returns an empty shell. A browser MCP server can load such a page. It is meant for one purpose: when you give only the URL of a posting, the plugin should still be able to read it.

Browser route switched off until redirects are shown

This page covers the intake of a posting from a URL (spec/job-application/stellenanzeige-url/en.md), implemented in job-application-posting-analysis. The route through the open-webSearch server works. The browser route stays switched off in practice: it needs a server that is shown to report the final URL of the loaded page, and none is (see "Limits you should know"). The plugin ships no .mcp.json. The rules all optional servers share (consent, one route per source, budget, no evasion) are on Optional MCP servers. The server is an independent project, not affiliated with this plugin. No skill requires one: without a server you paste the posting text, as described in Getting started.

Every statement about the server is the project's own, as its README says it. The README was read on 24 October 2026 on its main branch, not at a tagged version. The latest npm version on that day was @playwright/mcp 0.0.83. The plugin has not run the server and does not confirm these statements. Check the README for anything newer.

How the plugin uses a URL

  1. You give a URL. Tracking parameters such as utm_source are removed first.
  2. The plugin first fetches the page without a browser through the open-webSearch server.
  3. Only if that returns a page that is just a script shell does it offer the browser, name the server and the risk, and wait for your yes. A removed posting, a log-in wall or a bot check never leads to the browser; you are asked to paste the text. While no server is shown to report the final URL, this step does not happen and you are asked to paste the text.
  4. The browser opens the one URL, reads the page text and closes. It does not click, type, fill forms, upload or download, and it keeps no profile, cookies or logins.
  5. You see the extracted posting and its fields before anything is written to your Wurzel. A URL that lists several postings is not a posting: you name one entry. linkedin.com links and internal addresses are refused before any request.

A page behind a log-in or a CAPTCHA is not read, and no workaround is tried. A linkedin.com URL is never loaded this way; see LinkedIn access. The research of a firm (levels 1 and 2) never uses a browser.

The server

Playwright MCP
Project microsoft/playwright-mcp
Package, licence @playwright/mcp, Apache-2.0
Needs Node.js 18 or newer
Browser Starts its own browser
Profile and logins --isolated keeps the profile in memory and saves nothing
Headless --headless

The plugin reads a posting without your logins and without your history. Use only a server that can run in an isolated, non-persistent browser context; a server that works in your own browser session does not fit this intake.

Set up Playwright MCP

The README of the project is the authority for current steps and options. For Claude Code:

claude mcp add playwright -- npx @playwright/mcp@<version> --isolated --headless --block-service-workers

Replace <version> with a version number you have checked. The README shows @latest, which runs new code on every start without your review. The version checked for this page is 0.0.83.

Option What it does, per the README
--isolated Keeps the browser profile in memory and does not save it to disk. Without it, a persistent profile is kept in a cache folder
--headless Runs without a visible window. The default is a visible window
--block-service-workers Blocks service workers

Limit what the server can do

The server offers many more tools than the intake needs. The plugin uses only browser_navigate, browser_snapshot and browser_close. Deny the other tools that act on a page in your Claude Code permission settings, for example in .claude/settings.local.json:

{
  "permissions": {
    "deny": [
      "mcp__playwright__browser_click",
      "mcp__playwright__browser_drag",
      "mcp__playwright__browser_drop",
      "mcp__playwright__browser_evaluate",
      "mcp__playwright__browser_file_upload",
      "mcp__playwright__browser_fill_form",
      "mcp__playwright__browser_hover",
      "mcp__playwright__browser_press_key",
      "mcp__playwright__browser_run_code_unsafe",
      "mcp__playwright__browser_select_option",
      "mcp__playwright__browser_type"
    ]
  }
}

Tool names follow the README. Open /mcp after the setup and check the list of tools of your installed version; a name that is missing from the list above is not denied. The README says that browser_run_code_unsafe is the riskiest one: it runs arbitrary code in the server process.

Limits you should know

  • Not a security boundary. The README says so for Playwright MCP as a whole. Its option --allowed-origins "does not affect redirects", so it cannot stop a page from sending the browser somewhere else.
  • Redirects are not confirmed. The plugin must check where a page ended up, because a posting link can redirect to a page it must not read (RC-5.2 rule 9). The README does not say that a tool reports the final URL. Until a run of the installed version shows it, the browser route stays switched off in practice. This is an open point in the specification (UE-12).
  • Scripts run on your machine. A browser executes the scripts of whatever site the URL points to. That is why the plugin asks for your yes before it opens one.
  • Terms of the site. Some sites forbid automated reading. You check their terms; the plugin does not claim a fetch is allowed. A blocked page is reported and skipped, and no proxy or other workaround is used.
  • Posting text is data. Text on the page that tries to give instructions is ignored.
  • Personal data. What the browser reads becomes part of the Claude session and goes to the model provider. The plugin stores only the posting text and the fields it extracts.

Without a browser server

Paste the posting text into job-application-solution. Every skill works that way.