Skip to content

Web search

Company research needs two simple things from the web: finding a page and reading a page you already know, such as an imprint, a career page or a press release. The open-source open-webSearch MCP server gives Claude both, runs on your own machine and needs no API key.

The preferred route for web search, but not part of this plugin

With the server connected, job-application-posting-analysis and job-application-fit-check search the web and fetch public pages through it. Without it, they use the built-in tools of your client under the same query rules, and the record names the route that was used. This is the working rule of the research procedure (spec/job-application/recherche/en.md, RC-5.2 rules 8 and 9, RC-5.5 rules 4 to 11). No skill requires the server. The plugin installs nothing and ships no .mcp.json: you add the configuration below yourself. The server is an independent project, not affiliated with this plugin. The rules all optional servers share are on Optional MCP servers.

What the server does

The research uses three of the server's six tools. The other three are switched off in the configuration below.

Tool Use in the research
search Find a URL or a lead for a firm: the exact firm name, optionally the Sitz, and one allowed generic term such as Impressum or Pressemitteilung (RC-5.5)
fetchWebContent with renderMode: request Read one known public page as Markdown: imprint, career page, press, privacy notice (RC-5.2)
fetchGithubReadme Read the README of a firm's public repository

fetchWebContent also knows the modes auto (its default) and browser. The research always asks for request, so no browser renders a page.

Why this route comes first: its engines are documented and limited to a short list, the budget and the queries follow the plugin's rules, and this page can say what leaves your machine. With whatever search tool a session happens to have, that depends on the client, the region and the account.

Set it up

You need Node.js, because the server starts through npx. The server's own README is the authority for current options. Add this entry to the mcpServers section of your MCP configuration:

{
  "mcpServers": {
    "websearch": {
      "command": "npx",
      "args": ["open-websearch@2.2.0"],
      "env": {
        "MODE": "stdio",
        "DEFAULT_SEARCH_ENGINE": "duckduckgo",
        "ALLOWED_SEARCH_ENGINES": "duckduckgo,startpage,brave",
        "SEARCH_MODE": "request",
        "USE_PROXY": "false",
        "MCP_TOOL_FETCH_CSDN_NAME": "<disabled>",
        "MCP_TOOL_FETCH_LINUXDO_NAME": "<disabled>",
        "MCP_TOOL_FETCH_JUEJIN_NAME": "<disabled>"
      }
    }
  }
}

Version checked

Version 2.2.0 (last checked in October 2026). The variable names and defaults below were checked against the README of version 2.2.0. Before you change the pinned version, read the new README again and compare the variable names.

Setting Why
MODE=stdio The default is both, which also opens an HTTP server next to stdio. With stdio, no port is open. CORS stays off (ENABLE_CORS is false by default)
DEFAULT_SEARCH_ENGINE=duckduckgo The default engine would be bing
ALLOWED_SEARCH_ENGINES=duckduckgo,startpage,brave The default is empty, which allows every engine, among them several that are irrelevant here or Chinese-language. The research uses only these three
SEARCH_MODE=request The default is auto. According to the README, the setting currently affects only Bing. It states your intent: no browser
USE_PROXY=false This is the default. The plugin does not recommend a proxy to get around blocking
MCP_TOOL_FETCH_…_NAME=<disabled> According to the README, <disabled> switches a tool off. This disables the CSDN, Linux.do and Juejin fetch tools, so the tool list stays short and unambiguous

Type the value <disabled> literally, with the angle brackets. The README of version 2.2.0 documents this; an older release may not support it, which is one more reason to keep the pinned version. If /mcp still lists the three tools, check the spelling and the pinned version before you rely on the setting.

Engines. The server supports more engines, among them bing, baidu, hackernews and exa. Add bing only when you ask for it, for example because coverage of a German firm is too thin: append it to ALLOWED_SEARCH_ENGINES. The privacy properties of each engine are those of its operator. This page does not assert them.

Pinned version instead of @latest. The server's README shows npx open-websearch@latest. That runs new code on every start without your review. The configuration above names a version instead. Someone, usually you, reads the changes and checks the variable names before the pinned version moves. There is no automation for this. The open-websearch agent skill that the README offers is not used: the plugin has its own rules for queries and budgets.

To check that the server runs, open /mcp in Claude Code. The server websearch should be listed as connected. If it is not, the research uses the built-in tools instead.

Budget

The research keeps a budget per Vorgang. Calls run one after the other, never in bulk and never on a timer. You may lower the numbers.

Research level Searches Page fetches
1 at most 10 at most 15
2 at most 20 at most 30

The record in ## Recherche lists the route used and every search with engine, query and date (RC-9), so you can check afterwards what left your machine.

Limits and risks

  1. Queries leave your machine. Each query goes to the search engine. The query rules of the research apply unchanged (UA-3.2, RC-5.5): the exact firm name, optionally the Sitz, and one generic term from the allowed list. Never your profile, your Datengrundlage, your Wunschprofil or application text. What the engine operator does with a query is its own matter. This page makes no claim about it.
  2. Scraping and terms. The server reads engine result pages. Its README says the tool is "for personal use only", that you should comply with the terms of service of the corresponding engines, and it advises rate limiting. This is the server's statement, not legal advice, and this plugin does not confirm it. Using the server is your decision and your risk. The budget above limits how much you trigger.
  3. Results are fragile. Results depend on each engine's page layout and can break when it changes. A failed or blocked search leaves the items it served as nicht ermittelt, and the research goes on with other channels. A blocked engine is reported to you and skipped. No proxy is used to get around the block.
  4. A hit is a lead. A search result is a hint. A finding needs the page itself, read and dated, not the snippet (RC-5.5 rule 1).
  5. No browser. The research does not use Playwright or renderMode: browser. A page that is readable only with JavaScript is recorded as nicht ermittelt with the reason. You may paste the text yourself. The one exception is a posting URL that you give yourself; see Browser MCP.
  6. Not for LinkedIn and not for registers. A linkedin.com URL is never passed to the fetch tools (RC-5.4 rule 6). Use the LinkedIn server or paste the text. Handelsregister, Unternehmensregister and the insolvency notices are form portals and cannot be fetched this way.
  7. Fetched text is data. A page can contain instructions. Claude reads it as data and never follows it. According to the README, the initial and final URLs of fetchWebContent remain subject to public-network safety checks. The plugin passes only URLs from trusted places and never an internal address. The check of a redirect comes after the request: a redirect to linkedin.com is noticed only afterwards, and the text is then not used. The request to the redirect target has already been sent from your machine at that point. A built-in fetch tool that does not report the final URL is not used for pages at all. Most client fetch tools do not say so, so in practice only their search is available: the pages wait for the server or for text you paste, and the items stay open. Its search stays allowed.
  8. Everything read enters the session. Fetched pages become part of the Claude session and go to the model provider like any other text in it (RC-10).
  9. Supply chain. npx downloads and runs code from the npm registry. A pinned version limits this to code of one release, but does not replace a review. The licence is not repeated here; see the project's repository.
  10. No shipped configuration. Starting third-party code without a separate step of yours is not acceptable, so the plugin does not register the server. If you remove the entry, the research falls back to the built-in tools.