MCP for web search: giving Claude, Cursor and local models the internet

Q: Can I get web search for a local model?

Yes. In LM Studio, add a search MCP server to mcp.json and load a model that supports tool calling. The conversation stays on your computer; only the search queries leave it. See MCP in LM Studio .

The short version: a search MCP server gives a model two actions — find pages for a query, and read one of them in full. The model decides for itself when to search, and answers from current sources instead of from memory as of its training date. A remote server connects by URL, with nothing to install, to Claude, Cursor, LM Studio, Codex and OpenCode.

Why a model needs search when it "knows everything"

A model's knowledge ends at its training date, and in development and information work what matters is precisely the recent: a new version of a library, a changed API, yesterday's news, current prices. Without search the model either honestly says it does not know, or confidently hands you something out of date. With search it checks against the source and gives you the link. How MCP servers work in general, and how remote differs from local, is in the overview of MCP servers.

Two tools: find and read

Tool	What it takes	What it returns
`web-search`	A search query	The main result with a short summary, plus a list of other findings with links
`web-crawl`	A page address	The full text of the page — so the model reads the source, not a snippet

The pair works like this: the model searches, picks the most precise page out of the results — documentation, an article, a changelog — and reads it. That is why the answers land on substance instead of paraphrasing headlines from a results list.

How to ask so that search works well

Name what exactly you need and how recent: "find in the official documentation how, in version 5, you configure…", "what pricing changes were announced this month". The model passes that specificity into the query; a general question gets you a general results page.

How to connect it

The server address is https://twelver.app/api/mcp/v1/web-search. Tick it in the builder and the URLs and an mcp.json assemble themselves — and there are step-by-step instructions per client:

Claude — a custom connector in claude.ai, or claude mcp add --transport http in Claude Code;
Cursor — an entry in ~/.cursor/mcp.json;
LM Studio — web search for a local model, through mcp.json;
Codex and OpenCode — a section in the config and a key from an environment variable.

No keys: paste the URL into your client and sign in — Claude, Cursor and any other OAuth client.

Copy the connector URL

Paste it into your MCP client

Build your server

Check the tools — URLs and config update themselves.

Connector URLs

Web Search

https://twelver.ru/api/mcp/v1/web-search

Image Search

https://twelver.ru/api/mcp/v1/image-search

{
  "mcpServers": {
    "twelver-web-search": {
      "url": "https://twelver.ru/api/mcp/v1/web-search"
    },
    "twelver-image-search": {
      "url": "https://twelver.ru/api/mcp/v1/image-search"
    }
  }
}

The same mcp.json works in LM Studio, Windsurf, Cline, VS Code and any client that reads an mcp.json.

How this differs from Perplexity MCP

Search MCP servers split into two approaches, and the choice comes down to who is supposed to do the thinking about the results.

	An "answerer" server (Perplexity)	A search server
What it returns	A finished answer written by its own model, with sources	Results and page text — the conclusions are drawn by your model
Who reasons	The search engine's model, and then yours again	Only your client's model
When it is better	You want a short answer to a factual question	You need to read primary sources: documentation, articles, specifications
What it needs	A Perplexity API key and plan	The server URL; sign-in by OAuth or key

For coding agents the second approach is usually the more useful one: Claude or GPT inside Cursor reason better when they can see the original documentation rather than someone else's summary of it. Other options — Brave Search, Tavily, Exa — are collected in the shortlist of the best MCP servers.

Image search

A separate server, https://twelver.app/api/mcp/v1/image-search, finds existing images through the Yandex Images index — by description, with filters, and by the picture itself when you need to find its source or something similar. This is not generation: for new pictures there are the image generation servers.

What search over MCP costs

Search draws on the shared Twelver balance. The Basic plan at $8 a month is at least 580 searches, if you spend it on nothing else. Reading a page with web-crawl is billed separately. Figures as of 6 September 2026.

Frequently asked questions

Which MCP server should I choose for web search?

If the model needs to read primary sources — one that returns results plus the full text of pages: search plus page reading. If you want a short finished answer — an "answerer" server like Perplexity MCP. For coding agents the first is usually more useful: the model reasons over the original documentation.

How do I give Claude access to the internet through MCP?

Connect a remote search MCP server: in claude.ai through Settings → Connectors → Add custom connector, in Claude Code with claude mcp add --transport http. Once signed in, Claude calls search by itself whenever a question needs current data. In detail: MCP in Claude.

Can I get web search for a local model?

Yes. In LM Studio, add a search MCP server to mcp.json and load a model that supports tool calling. The conversation stays on your computer; only the search queries leave it. See MCP in LM Studio.

How is MCP search different from Perplexity MCP?

Perplexity MCP returns a finished answer written by the Perplexity model. A search MCP server returns results and page text, and the conclusions are drawn by your client's model. The first is handier for quick facts, the second for working with documentation and primary sources.

Is there an MCP server for image search?

Yes — a separate one, built on the Yandex Images index. It finds existing images by description or by an image you provide, which is what you want for locating a source or similar pictures. Creating new images is a different set of servers.

Try it yourself

Everything in this guide runs inside Twelver

One chat for text, images, video, music and voice — no separate services or subscriptions.

Open Twelver chat

Related pageMCP Toolsets

Rate your experience