How to Add a Remote MCP Server to LM Studio

How do I add a remote MCP server to LM Studio?

In a chat, open the right-hand sidebar, click the Program tab, then Install > Edit mcp.json. Add the server to the mcpServers object and save — LM Studio opens your browser for the OAuth consent screen automatically on save. You need v0.4.10 or later — MCP arrived in v0.3.17 but OAuth for MCP servers only landed in v0.4.10, so on an older build no consent screen ever appears — and, critically, a model that supports tool calling. That last requirement is not a formality: many local models don't support it, and a model that can't call tools will sit there ignoring a perfectly working server.

Requirements

• LM Studio v0.4.10 or later (free), with at least one model downloaded. MCP itself landed in v0.3.17, but OAuth for MCP servers shipped in v0.4.10 — and OAuth is what this guide uses, so an older build will follow every step and never prompt you to sign in
• A model that supports tool calling — check the model card; not every local model does
• A default browser set at the OS level, for the authorization step

Adding the server

1. In a chat, open the right-hand sidebar
2. Click the Program tab
3. Click Install > Edit mcp.json

Then add the server to the mcpServers object and save:

{
  "mcpServers": {
    "my-server": {
      "url": "https://mcp.example.com/mcp"
    }
  }
}

Authorization

When you save mcp.json, LM Studio opens your browser with the server's consent screen automatically. Sign in if prompted and approve. If the server supports dynamic client registration, there are no keys, client IDs, or secrets to paste.

If no browser window opens, re-save mcp.json to retrigger the flow, and check that a default browser is set at the OS level.

Approving tool calls

LM Studio asks for confirmation before each tool call. The first time your model calls one of the server's tools, a confirmation dialog appears — approve it. The same dialog lets you relax future confirmations for tools you trust.

This is worth knowing before you conclude something is broken: a first tool call that appears to hang may simply be waiting on a dialog.

The two failures specific to local models

Both of these are about the model, not the server.

The model never calls the tools. Confirm your model supports tool calling. Many local models don't, or do it unreliably, and the model card is the place to check. There is no config change that fixes this — you need a different model.

Responses got worse after connecting. This one surprises people. Smaller local models can struggle when a large tool list lands in their context: the tools crowd the window and the model's ordinary answers degrade. Try a larger model, or disable the tools you don't need for the current chat.

That second failure is the honest trade-off of connecting a rich server to a small local model, and it is better to know it up front than to conclude the server is at fault.

What else goes wrong

• No browser window for authorization. Re-save mcp.json; check the OS default browser.
• The connection fails immediately. Check the URL against exactly what the server publishes, including the path and any trailing slash.

A worked example

LM Studio runs open-source models locally — Llama, Qwen, GLM, Kimi, DeepSeek, gpt-oss, anything in GGUF or MLX format. What those models lack is persistence: the model is on your machine, and nothing carries between sessions.

Tempreon is a hosted MCP server that supplies exactly that missing half:

{
  "mcpServers": {
    "tempreon": {
      "url": "https://api.tempreon.com/functions/v1/tempreon-mcp/mcp"
    }
  }
}

Save, approve in the browser, approve the first tool call, and a fully local model has a persistent identity — the model runs entirely on your machine, and the memory comes from the server. There is also a one-click install link and the account-holder walkthrough in the LM Studio setup doc.