webstractor

How to Add Web Search to LM Studio with Webstractor MCP

LM Studio can run the model locally while a remote MCP tool supplies current public-web results. This separation is useful when you want to change local models without rebuilding search integration code. Webstractor returns compact result titles, snippets, and source URLs, leaving the model to compare and synthesize the evidence.

Connect one remote MCP server in mcp.json.

In LM Studio, open the Program tab, choose Install, then Edit mcp.json. Add Webstractor as a remote server, save the file, and use a local model with reliable tool calling. The search_web tool can then return ordered, source-linked results without changing your model provider.

MCP configuration
{
  "mcpServers": {
    "webstractor": {
      "url": "https://webstractor.com/mcp"
    }
  }
}

What you need

  • LM Studio 0.3.17 or newer
  • A downloaded model with tool-use support
  • Network access to webstractor.com

Connect Webstractor to LM Studio

01

Load a tool-capable model

Download and load a model whose model card or prompt template supports tool calling. Small instruction-only models may see the tool definitions yet never invoke them.

02

Open the MCP configuration

In the right sidebar, switch to Program, select Install, and click Edit mcp.json. Preserve any servers already present in the mcpServers object.

03

Add the hosted Webstractor server

Paste this named server entry and save. LM Studio supports remote HTTP MCP servers directly, so no local Node or Python bridge is required.

mcp.json
{
  "mcpServers": {
    "webstractor": {
      "url": "https://webstractor.com/mcp"
    }
  }
}
04

Confirm the tools loaded

Return to the Program panel and inspect Webstractor. You should see search_web alongside the other search and extraction tools.

05

Run a source-first test

Prompt: “Use search_web to find the current LM Studio MCP documentation. Return five results, then summarize only what those sources support and include their URLs.” A visible tool call plus linked results confirms the connection.

Tool calling matters more than model size alone

MCP only makes tools available; the selected model decides whether and how to use them. If a model answers from memory, explicitly name search_web in the prompt and test a stronger tool-use model before changing the server configuration.

Keep the requested result count modest. Compact result sets consume less context and make it easier for a local model to compare sources.

Search first, extract second

Use search_web when you do not yet know the right page. Once LM Studio identifies a promising source, ask it to call extract_public_url for the full readable page rather than treating a search snippet as complete evidence.

  • Discovery: search_web with a precise query
  • Selection: choose the most relevant canonical source
  • Reading: extract_public_url on that source
  • Answering: cite the original source URL

If LM Studio does not call the search tool

Check that the JSON is valid and nested under one mcpServers object. Reload the server after edits, verify the model supports tools, and start with a direct prompt that names search_web.

If the server appears but a long query times out, retry with a concise query and fewer results. Do not add an SSE URL: Webstractor uses the single Streamable HTTP endpoint shown above.

01

What you can extract

  • Ordered public-web results with titles, snippets, and URLs
  • Up to 10 sources from one search request
  • Optional language, country, and site controls
  • Markdown for reading or schema-v1 JSON for downstream logic
02

Where normalized data helps

  • Ground a local model with current technical documentation
  • Find primary sources before drafting a research answer
  • Discover public pages for a local RAG collection
  • Restrict discovery to one documentation hostname

Public data only

  • LM Studio still needs a model that handles tool calls well.
  • Web results are source material, not a guarantee that every claim is correct.
  • Only send public queries; do not put credentials or private data into search terms.

webstractor.com does not bypass CAPTCHAs, login walls, paywalls, access controls, or regional restrictions. Review the source’s terms and applicable law before collecting or reusing data.

LM Studio and Webstractor FAQ

Does adding Webstractor make LM Studio fully offline?

No. Model inference can stay local, but current web search requires an outbound connection from LM Studio to the hosted MCP endpoint and from that service to public sources.

Do I need to install an npm package?

No. LM Studio supports remote MCP servers, so the URL entry is enough.

Why does the model answer without searching?

The model controls tool selection. Use a tool-capable model and explicitly ask it to use search_web and cite returned URLs.

Turn a public URL into useful context.

Connect Webstractor MCP Read the Webstractor MCP docs