MCP in LM Studio: web search and image generation for a local model

The short version: from version 0.3.17 LM Studio can connect MCP servers through an mcp.json file. A remote server is added with two fields — url and headers with a key. After that, a local model can search the web, read pages and generate images, none of which it can do on its own.

Why a local model needs MCP

A model in LM Studio knows the world as of its training date and can only write text. Ask it about yesterday's news and it will either admit it does not know or, worse, invent an answer. Ask it to draw a logo and it will describe one in words: LM Studio is built for language models, and language models do not draw.

MCP closes both gaps without touching the model itself. You connect tools, and the model starts deciding for itself when to go and search and when to call a generator. The heavy lifting happens on the remote server: image generation needs no second graphics card and no wrestling with separate diffusion models.

Step 1. A model that can call tools

MCP only works with models trained to call functions (tool use, function calling). In LM Studio's catalogue those models carry a corresponding badge. Modern mid-sized models do the job — the Qwen3 and gpt-oss families, for instance. Very small models technically support calls but get confused about parameters and reach for tools at the wrong moments.

The second requirement is context. Tool descriptions take up room in every request, and a search result or a fetched page takes up more. Set the context length to at least 8–16 thousand tokens in the model's load settings, or tool responses will be truncated.

Step 2. Get a key

LM Studio passes the server a static header, so sign-in here is by key rather than through an OAuth window. The button below issues one, no registration:

Step 3. Write the servers into mcp.json

In LM Studio, open the Program tab in the right-hand panel and choose Install → Edit mcp.json. The file opens in the built-in editor. Add the servers inside the mcpServers block:

{
  "mcpServers": {
    "twelver-web-search": {
      "url": "https://twelver.app/api/mcp/v1/web-search",
      "headers": {
        "Authorization": "Bearer sk-<your-key>"
      }
    },
    "twelver-image": {
      "url": "https://twelver.app/api/mcp/v1/image-generation",
      "headers": {
        "Authorization": "Bearer sk-<your-key>"
      }
    }
  }
}

Save the file. The servers appear in the list on the Program tab, where a toggle switches each on and off without removing it from the config.

Step 4. Check it in a chat

Load a tool-capable model, open a new chat and ask for something the model cannot possibly know:

• "Search the web for the latest version of LM Studio and summarise what's new in it" — the model will call web-search, and then most likely web-crawl on the link it found.
• "Draw a cover for a podcast about coffee, square" — the model will call text-to-image and return a link to the picture.

Before a call, LM Studio shows which tool the model wants to run and with which parameters — you can allow or decline. For search it is convenient to always allow; for generation, confirm by hand, since it spends your balance.

What comes back from a generation

The generation server returns not the picture itself but a link to a file — the model puts it in its answer, and the image opens in the browser when you click it. That keeps the result from bloating the local model's context with megabytes of data, and the file stays reachable after the chat is closed.

On the starting tokens the picture arrives with a small wordmark on it; after the first purchase files are available in the original and can be downloaded. Everything about choosing engines — Nano Banana, Seedream, video through Seedance and Hailuo — is in image and video generation over MCP; the finer points of search are in MCP for web search.

If it does not work

The same server connects to Claude and Cursor — and the big picture, what MCP is and what it is for, is in the overview of MCP servers.