Skip to main content
PromptQuorum
Home/Power Local LLM/web-search-mcp Review: Web Search Tool Access for Local LLMs
Local AI Agents & Tool Use

web-search-mcp Review: Web Search Tool Access for Local LLMs

Β·8 min readΒ·By Hans Kuepper Β· Founder of PromptQuorum, multi-model AI dispatch tool Β· PromptQuorum

web-search-mcp is a free, open-source, self-hosted Model Context Protocol (MCP) server that gives a local LLM in a client like LM Studio or LibreChat a working web-search tool, without requiring a search API key. It runs on your own machine as a small Node.js process, but each search it performs is a live request to a public search engine (prioritized Bing, then Brave, then DuckDuckGo) β€” the tool is self-hosted, but the searches it returns are not offline or private in the sense of never leaving your machine.

web-search-mcp (github.com/mrkrsl/web-search-mcp) is a free, open-source, self-hosted Model Context Protocol (MCP) server that gives a local LLM β€” in tools like LM Studio or LibreChat β€” the ability to call out and run real web searches and fetch page content, without requiring a search-provider API key. It is a tool-use bridge, not an offline search index: it works by sending live requests to public search engines over the internet each time the model calls it. This review covers what it does, how it works, and who it fits.

Key Takeaways

  • web-search-mcp (github.com/mrkrsl/web-search-mcp) is a free, open-source, self-hosted MCP server that adds web search and page-content tools to MCP-compatible AI clients
  • Built by mrkrsl in TypeScript; MIT licensed, confirmed via the repository's LICENSE file
  • No search-provider API key required β€” uses Playwright browser automation with a fast HTTP-request fallback, prioritizing Bing, then Brave, then DuckDuckGo for reliability
  • This is a self-hosted tool, not a private/offline one: it fetches live results from external search engines over the internet every time it runs
  • Exposes three MCP tools to the connected model: full-web-search (search plus extracted page content), get-web-search-summaries (search results without full content), and get-single-web-page-content (content extraction from one given URL)
  • Documented setup steps for LM Studio and LibreChat, including a Docker path for LibreChat; usable with any client that speaks MCP
  • This directory entry lists it as a companion tool for the Local LLM Software Directory under agent-frameworks / tool-use servers

πŸ“ In One Sentence

web-search-mcp is a free, open-source (MIT), self-hosted Model Context Protocol server built in TypeScript that gives local LLMs in clients like LM Studio and LibreChat working web-search and page-fetch tools, using live Bing/Brave/DuckDuckGo requests instead of a paid search API.

πŸ’¬ In Plain Terms

web-search-mcp is a small program you run on your own computer that plugs into an MCP-compatible AI app (like LM Studio) and gives the AI model the ability to actually search the web and read pages, the way a human would, instead of only knowing what it learned during training. It is self-hosted, but it is not private browsing β€” the search requests it makes leave your machine and go to real search engines on the internet, the same as opening a browser.

πŸ“ŒNote: This review is based on the project's own GitHub repository, its README, and its documentation for LM Studio/LibreChat setup. It does not claim PromptQuorum ran its own hands-on benchmark of search latency or accuracy, and it does not describe web-search-mcp as private or offline β€” its core function is fetching live results from public search engines over the internet.

What Is web-search-mcp?

web-search-mcp is a free, open-source, self-hosted Model Context Protocol (MCP) server that gives an AI model the ability to run real web searches and fetch page content, called on demand by the model itself. It is built by mrkrsl and distributed on GitHub. The name is shared by several unrelated MCP projects; this review covers specifically github.com/mrkrsl/web-search-mcp, whose documentation explicitly targets local-LLM setups like LM Studio.

  • Product type: a self-hosted MCP server process, not a standalone desktop app or a hosted API β€” you run it yourself alongside an MCP-compatible client
  • Creator: GitHub user mrkrsl
  • License: MIT, confirmed via the repository's LICENSE file
  • Language/runtime: TypeScript, requiring Node.js 18.0.0+ and npm 8.0.0+ per the project's stated requirements
  • Funding: no funding round, investor, or commercial backing was found for this review β€” treat it as an independently maintained, community open-source project
  • Scale: roughly 1,100+ GitHub stars as of this review (GitHub's UI shows the rounded "1.1k" badge rather than an exact count)

What Does web-search-mcp Actually Do?

web-search-mcp runs as a local MCP server process that an MCP-compatible AI client connects to, and it exposes three tools the connected model can call: a full web search with extracted page content, a lighter search-summaries-only mode, and single-page content extraction from a given URL.

  • full-web-search: runs a search query and returns extracted content from the resulting pages, giving the model enough text to answer questions grounded in current web content
  • get-web-search-summaries: runs a search but returns only result summaries/snippets, not full page content β€” faster and lighter for cases where the model just needs an overview
  • get-single-web-page-content: fetches and extracts readable content from one specific URL the model already has, without running a new search
  • Search backend strategy: per the project's own documentation, it "prioritises Bing > Brave > DuckDuckGo for optimal reliability and performance," falling back down the list if one backend fails
  • "Smart Request Strategy": switches between Playwright-driven browser automation (for pages that need JavaScript rendering) and faster, lighter axios-based HTTP requests depending on the target site, per the project's README
  • No search-provider API key needed β€” this is the main practical difference from API-key-gated search tools (e.g., Bing Search API, Brave Search API, SerpAPI), which require a paid account and key to use programmatically
  • Configurable via environment variables covering content-length limits, request timeouts, concurrent-browser limits, and browser engine choice, per the project's documentation

Platform, Pricing, and Licensing

Platform

What web-search-mcp states:
A self-hosted Node.js/TypeScript server process, requiring Node.js 18+ and npm 8+. Documented as working on macOS, Windows, and Linux.

Cost

What web-search-mcp states:
Free and open source. No search-provider API key or paid tier is required to use it.

Licensing

What web-search-mcp states:
MIT license, confirmed via the repository LICENSE file.

Locality

What web-search-mcp states:
Hybrid: the server process itself is self-hosted, but each search or page fetch is a live network request to an external search engine or website β€” this is not an offline or fully local tool.

Install method

What web-search-mcp states:
Download or clone the repository, then run npm install, npx playwright install, and npm run build, per the project README.

Verify current install steps on github.com/mrkrsl/web-search-mcp before running any command, since setup instructions can change between releases.

web-search-mcp vs. Built-In Browsing

Compared to a chat app's own built-in web-search or browsing feature (where one exists), web-search-mcp trades convenience for control: you self-host it, choose which client connects to it, and are not limited to whatever single provider a closed app has integrated.

Setup

web-search-mcp:
Manual: clone, npm install, configure mcp.json
Built-in app browsing:
Usually a toggle in settings, no setup

Works with local LLMs

web-search-mcp:
Yes β€” documented for LM Studio, LibreChat, any MCP client
Built-in app browsing:
Varies; often tied to one hosted-model provider

Search backend

web-search-mcp:
Bing / Brave / DuckDuckGo, self-selected fallback order
Built-in app browsing:
Fixed, usually undisclosed provider

API key needed

web-search-mcp:
No
Built-in app browsing:
No (cost is bundled into the app/plan)

Runs fully offline

web-search-mcp:
No β€” hybrid, live external search each call
Built-in app browsing:
No β€” same constraint

Both approaches call out to the open internet; neither is an offline search tool. The distinction is who hosts the bridge and which client/model it works with.

Who Should Use web-search-mcp?

web-search-mcp fits developers running a local LLM through an MCP-compatible client who want to add real-time web lookups without paying for a search API.

What web-search-mcp Is Not Good For

web-search-mcp is not a good fit if you need fully offline operation, a zero-configuration setup, or search reliability backed by a paid, contracted API.

  • Not offline or private in the sense of "never leaves your machine" β€” every search or page fetch is a live request to an external search engine or website over the internet; only the server process is self-hosted
  • Not a zero-setup tool β€” it requires Node.js, npm, a Playwright browser install, and manual configuration of an MCP client's mcp.json file
  • Not backed by a paid search API's uptime or rate-limit guarantees β€” automating public search-engine pages can be more fragile under heavy or automated use than a contracted, paid search API
  • Not a general-purpose plugin β€” it only works with clients that implement the Model Context Protocol, not arbitrary chat apps or browser extensions
  • Not independently benchmarked by PromptQuorum for search latency, result accuracy, or uptime β€” this review reflects the project's own documentation, not hands-on measurement

Common Mistakes When Evaluating web-search-mcp

Most confusion about web-search-mcp comes from mixing up "self-hosted" with "offline," or assuming it is the only project with this name.

Competitors and Alternatives

Within local AI agent tooling and MCP-connected tool servers, web-search-mcp is most comparable to other self-hosted agent and research tools that add live information retrieval to a local LLM.

Tool
Best known for
Link
nanobotSelf-hosted personal AI agent with persistent memory and MCP tool supportnanobot Review
Local Deep ResearchSelf-hosted, multi-step AI research assistant that synthesizes web and local sourcesLocal Deep Research Review
MulticaLocal AI agent and tool-use platform for connecting models to external toolsMultica Review

This list reflects tools commonly compared to web-search-mcp within the local AI agents / tool-use segment, not an independent PromptQuorum ranking β€” verify each tool's current feature set before choosing.

Frequently Asked Questions

What is web-search-mcp?

web-search-mcp (github.com/mrkrsl/web-search-mcp) is a free, open-source (MIT), self-hosted MCP server that gives AI models in MCP-compatible clients real web-search and page-content-fetch tools.

Is web-search-mcp free?

Yes, it is free and open source under the MIT license. There is no paid tier and no search-provider API key required.

Does web-search-mcp work fully offline?

No. The server process is self-hosted, but every search or page fetch it performs is a live request to an external search engine (Bing, Brave, or DuckDuckGo) or website over the internet.

Do I need a search API key to use it?

No. It uses browser automation (Playwright) with an HTTP-request fallback against public search engines rather than a paid search-provider API, so no API key is required.

Which AI clients does web-search-mcp work with?

Any client that implements the Model Context Protocol (MCP). The project documents specific setup steps for LM Studio and LibreChat.

How do I install web-search-mcp?

Download or clone the repository from GitHub, then run npm install, npx playwright install, and npm run build, per the project README. It requires Node.js 18+ and npm 8+.

What search engines does it use?

Per the project's own documentation, it prioritizes Bing, then Brave, then DuckDuckGo, falling back down the list if one backend fails.

What tools does it expose to the connected model?

Three MCP tools: full-web-search (search plus extracted page content), get-web-search-summaries (results without full content), and get-single-web-page-content (content extraction from one URL).

Who created web-search-mcp?

GitHub user mrkrsl created and maintains web-search-mcp.

Are there other projects also named "web-search-mcp"?

Yes. Several unrelated GitHub repositories share the same generic name. This review covers specifically github.com/mrkrsl/web-search-mcp, chosen because its documentation explicitly targets local-LLM setups.

Sources

← Back to Power Local LLM