LM Studio MCP, Plugins and Agent Mode (2026): mcp.json Setup, Best Tool Models and Bionic Explained

LM Studio started as a simple way to chat with local models. In 2026 it can also use tools: read files, search the web, drive a browser and edit code, all through MCP servers, plugins and the new Bionic agent app.

The catch is that these three things are easy to mix up, and a badly set up mcp.json silently breaks every tool at once.

This guide explains which one you need, gives copy-paste mcp.json examples for Windows, Mac and Linux, picks models that actually call tools, and fixes the common failures.

MCP vs Plugins vs Bionic: Which Do You Need?

Before You Start

• LM Studio 0.3.17 or newer. Current builds are in the 0.4.x line. Update from the app menu.
• Node.js 18+ for the many MCP servers that start with npx.
• A tool-capable model (hammer icon in the model list).
• A sandbox folder for file tools, so the model never touches your whole drive.

New to LM Studio itself? My Ollama vs LM Studio vs llama.cpp comparison explains where it fits.

How to Add an MCP Server in LM Studio

1. Open a chat and show the right-hand sidebar.
2. Go to the Program tab, then click Install > Edit mcp.json. The file opens in LM Studio’s editor.
3. Add your server inside mcpServers (examples below) and save. Servers load on save.
4. Turn the server on for this chat in the tools/integrations list.
5. Load a tool-capable model with at least 16K context.
6. Ask something that needs the tool, for example “list the files in my sandbox folder”.
7. Approve the call in the confirmation dialog. You can edit the arguments and choose allow once or always.

Tip: some sites offer an “Add to LM Studio” button. It uses an lmstudio://add_mcp link and writes the entry for you.

Warning: the file must have exactly one top-level mcpServers object. Pasting a second wrapper, or leaving a trailing comma, makes the JSON invalid and no servers load.

Copy-Paste mcp.json Examples

Filesystem (read and write one folder)

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/ai-sandbox"]
    }
  }
}

On Windows, if npx will not start, wrap it with cmd and use forward slashes in the path:

"filesystem": {
  "command": "cmd",
  "args": ["/c", "npx", "-y", "@modelcontextprotocol/server-filesystem", "C:/Users/you/ai-sandbox"]
}

Playwright (control a browser)

"playwright": {
  "command": "npx",
  "args": ["@playwright/mcp@latest", "--headless"]
}

Drop --headless to watch the browser window. Add "--browser", "chrome" (or firefox, webkit, msedge) to pick an engine.

Hugging Face (remote server)

"hf-mcp-server": {
  "url": "https://huggingface.co/mcp",
  "headers": { "Authorization": "Bearer <YOUR_HF_TOKEN>" }
}

GitHub (remote server)

The old @modelcontextprotocol/server-github npm package is archived. Use GitHub’s hosted server at https://api.githubcopilot.com/mcp/ with a personal access token in the Authorization header, in the same url + headers form as above.

Web search

LM Studio has no built-in web search. Add a search provider’s MCP server (Brave, Exa or Tavily, each with its own API key) or a simple fetch server. For more server ideas, see my list of the best MCP servers in 2026.

Best Models for Tool Calling by Memory

Tool use fails more often because of the model than because of MCP. These picks come from LM Studio-focused guides; sizes are approximate Q4 figures.

For more options by card size, see my best local models by VRAM ranking (the same GGUF files work in LM Studio) and how much VRAM you need.

Note: every enabled server adds its tool list to the prompt. Enable only the servers a chat needs, or a small context fills up before the model even answers.

LM Studio Plugins Explained

Plugins are LM Studio’s own extension system, written in TypeScript (Python support is in development). They are still marked beta.

• Tools providers: give the model new functions, like MCP but built for LM Studio.
• Prompt preprocessors: change your message before the model sees it, for example adding retrieved documents.
• Generators: replace the model’s output step with your own code.
• Custom configurations: add settings for the above.

1. Open the plugin’s page on the LM Studio Hub.
2. Click “Run in LM Studio”.
3. Restart LM Studio and switch the plugin on in the Tools panel.

Building your own? The lms CLI has lms clone owner/name, lms dev for a live dev loop and lms push to publish. The lmstudio-js and lmstudio-python SDKs are MIT-licensed.

LM Studio Bionic: The “Agent Mode” App

When people search “LM Studio agent mode”, they usually mean Bionic. LM Studio announced it on 16 July 2026 as “the AI agent for open models”. It is a separate desktop app, released as a preview.

• Code projects: link a local folder and let a model investigate, edit or debug it, with inline diffs and agentic code search.
• Work projects: create and edit PDFs, decks and spreadsheets in a sandbox, with automatic checkpoints you can roll back.
• Voice keyboard: system-wide dictation with Mistral’s Voxtral, transcribed on-device.
• LM Link: connect to a more powerful machine of yours over an encrypted mesh.

Tip: to be sure Bionic stays local, pick only chip-icon models, check lms ps shows the model loaded, and watch your GPU usage during a reply. An idle GPU means the request went to the cloud.

Using MCP Through the LM Studio API

LM Studio’s local server (OpenAI-compatible, port 1234 by default) can also use your MCP servers. Its /api/v1/chat endpoint accepts an integrations field and an allowed_tools list. Both are off by default, so a script cannot use your tools unless you enable them on purpose.

If you want a coding agent in VS Code instead, my Cline + local LLM guide shows the setup with a local server.

Safety Rules for Tools on Your PC

• Only install trusted servers. LM Studio’s own docs warn that MCP servers can run code, read files and use your network.
• Sandbox file access to one folder, never your home directory or a whole drive.
• Keep confirmation on for anything that writes, posts or deletes. “Always allow” is fine for read-only tools.
• Watch for prompt injection. A web page the model reads can contain instructions. Do not let a browsing session also have write access to important files.
• Keep tokens out of shared files. mcp.json holds API keys in plain text.

Troubleshooting

No MCP servers load at all

Almost always invalid JSON. Check for a missing brace, a trailing comma or two mcpServers blocks. Paste the file into any JSON validator.

Server fails to start on Windows

Install Node.js 18+ and restart LM Studio so it sees the new PATH. If it still fails, use the "command": "cmd" with "/c", "npx" form shown above.

The model ignores tools or loops

Raise the context length when loading the model, turn off servers you do not need, and use a model with native tool support (hammer icon). Small 7-8B models struggle with many tools.

Context overflow after a tool call

Servers built for big cloud models can return huge results, like full page snapshots. Use a larger context, a lighter server, or ask for smaller results (“first 20 lines only”).

GitHub server errors or “package deprecated”

You are on the archived npm package. Switch to the remote server at https://api.githubcopilot.com/mcp/ with a personal access token.

Bionic will not install on Linux

There is no Linux build yet. Use LM Studio with MCP, or a VS Code agent like Cline with a local server.

My Take

MCP is the part of LM Studio I would set up first. A filesystem server on a sandbox folder plus one search server turns a local model from a chatbot into something that can actually look things up and work with your files.

Bionic is promising, but treat it as a preview and keep an eye on the cloud icon if privacy is why you went local in the first place.

Prefer Ollama? You can pair it with Claude Code instead; see my Claude Code Router guide.

FAQ

Does LM Studio support MCP?

Yes, since version 0.3.17 (June 2025). LM Studio acts as an MCP host for local servers (command and args) and remote servers (url and headers).

Where is mcp.json in LM Studio?

At ~/.lmstudio/mcp.json on Mac and Linux, or %USERPROFILE%\.lmstudio\mcp.json on Windows. Open it from Program > Install > Edit mcp.json.

Does LM Studio have an agent mode?

Not as a switch in the chat app. Tools work in normal chats through MCP and plugins. LM Studio’s full agent is Bionic, a separate app released in preview in July 2026.

Is LM Studio Bionic free?

The app and local models are free. Cloud models run on pay-as-you-go credits, and a “Bionic Pass” subscription was listed as coming soon.

Why won’t my model use MCP tools?

Usually the model lacks tool training, the context is too small, or the server is switched off for that chat. Use a hammer-icon model with 16-32K context.

Does LM Studio have web search?

Not built in. Add a search MCP server such as Brave, Exa or Tavily with your own API key.

Are LM Studio plugins the same as MCP servers?

No. Plugins are LM Studio’s own TypeScript add-ons from the LM Studio Hub. MCP servers are a cross-app standard that also works in Claude, Cursor and VS Code.

Sources

• LM Studio docs: Use MCP Servers
• LM Studio blog: MCP in LM Studio (0.3.17)
• 9to5Mac: LM Studio launches Bionic
• RunAIHome: LM Studio Bionic hardware guide
• AI Agents: LM Studio plugins and MCP
• Local AI Master: LM Studio MCP setup