LM Studio Copilot Expansion by MyBestTools

Run local LM Studio models inside VS Code Copilot Chat with streaming responses and tool calling.

This extension uses the LM Studio HTTP API only. Whether your LM Studio server runs on the same machine or on a remote host, the connection flow is the same: point the extension at the server URL and use the models exposed by that API.

Install LM Studio Copilot Expansion by MyBestTools from the VS Code Marketplace when this edition is published.

What it does

Adds LM Studio models to the Copilot Chat model picker
Streams responses directly into VS Code chat
Uses the same API-based connection flow for both local and remote LM Studio servers

Requirements

VS Code 1.104.0 or later
.NET 8 runtime installed and on PATH — the extension's chat/tool logic runs in a bundled dotnet-hosted backend process
LM Studio server accessible at the configured URL (default: http://localhost:1234)
At least one chat model already exposed by that LM Studio server

Quick Start

Install this extension.
Start LM Studio and make sure its API server is reachable.
Load the model you want to use in LM Studio so it appears in the API model list.
Open Copilot Chat.
Pick an LM Studio model.
Send a prompt.

The extension does not start LM Studio, stop LM Studio, or load models on your behalf. It only talks to the configured API endpoint.

Connection Modes

Local and remote servers are handled the same way.

Examples:

Local LM Studio server: http://localhost:1234
Remote LM Studio server: http://your-server:1234
SSH tunnel to a remote server: http://localhost:1234

As long as the API endpoint responds and exposes a model list, the extension treats all of these the same. If you tunnel a remote LM Studio over SSH, its URL on your local machine will be http://localhost:1234, and the extension talks to it exactly the same as a real local server — no extra configuration needed.

Manual provider JSON fallback

If the VS Code GUI for adding a custom chat model is bugged, you can fall back to a manual provider definition. The customendpoint provider config is a top-level array of provider objects — paste an entry like this:

[
  {
    "name": "LM Studio",
    "vendor": "customendpoint",
    "apiKey": "lm-studio",
    "apiType": "chat-completions",
    "models": [
      {
        "id": "qwen/qwen3.6-27b",
        "name": "qwen/qwen3.6-27b",
        "url": "http://YOUR_TAILSCALE_IP:1234/v1/chat/completions",
        "toolCalling": true,
        "vision": false,
        "maxInputTokens": 80000,
        "maxOutputTokens": 16000,
        "streaming": true
      }
    ]
  }
]

Replace the model id, display name, and url with the values exposed by your LM Studio server. Set toolCalling and vision to match the capabilities of the model you are exposing.

Important Settings

Most users can leave the defaults alone. These are the settings that matter most:

lmstudio-copilot-expansion.serverUrl: LM Studio server URL (default: http://localhost:1234)
lmstudio-copilot-expansion.apiKey: API key used for LM Studio server authentication when required
lmstudio-copilot-expansion.autoRefreshModels: Refresh available models from the LM Studio API on startup
lmstudio-copilot-expansion.enableToolCalling: Enable tool calling for supported models
lmstudio-copilot-expansion.maxTools: Limit the number of tools exposed per request
lmstudio-copilot-expansion.logLevel: Controls output logging verbosity. Default is verbose; set to info, warning, error, or none.

Commands

LM Studio: Refresh Available Models
LM Studio: Review Suggested Model Config
LM Studio: Check Server Connection

Usage

Select an LM Studio model in Copilot Chat and start chatting. The extension only uses models currently exposed by the configured LM Studio API server.

Optional Features

Tool calling for supported models exposed by the LM Studio API

LM Studio Auto

Pick "LM Studio Auto" in the model picker instead of a specific model, and the extension routes each new conversation to whichever backing model fits it best — a quick model for short edits, a coding-tuned model for general work, a large-context model for planning/synthesis, a vision model when images are attached, and so on. It asks a currently loaded model to classify the request (falling back to a cheap heuristic if none is loaded or classification fails), then reuses that same pick for the rest of the conversation rather than re-classifying on every turn, since switching LM Studio models mid-chat means an unload/reload that can take real time. It shows which model it routed to as a short note at the start of the first response. It draws from taskTypeModels when configured for a profile, otherwise falls back to the best-scoring available model for that profile. Disable with lmstudio-copilot-expansion.enableAutoModel.

The active model is also made aware it can hand off the rest of the conversation itself, via a mbt_lmstudio_switch_model tool — only offered while "LM Studio Auto" is selected, never when you've pinned a specific model directly. If the model decides mid-conversation that a different profile fits better (e.g. it needs a bigger context window, or the remaining work is a much lighter follow-up), it can call the tool and the next response comes from the newly picked model.

Image Generation

Image generation is not built into this extension — configure it server-side as an MCP server instead, so it's available to any MCP-compatible client, not just this extension. See Local MCPs with LM Studio for setup instructions.

Troubleshooting

Models not appearing

Make sure LM Studio is running
Make sure the target model is loaded and visible through the LM Studio API server
Check lmstudio-copilot-expansion.serverUrl
Run LM Studio: Check Server Connection

Slow responses

LM Studio performance depends on your hardware and the model size
Consider using a smaller/faster model
Increase the timeout in settings if needed

Development

The extension itself (chat provider/tool registration, VS Code UI) is TypeScript in src/. All LM Studio API, tool-calling, and model-scoring logic lives in the C# backend under backend/, which the extension spawns as a child process and talks to over stdio (see backend/Program.cs and src/backend-process.ts for the protocol). Building the extension always rebuilds and republishes that backend alongside the TypeScript bundle.

# Install dependencies
npm install

# Run the repo validations
npm test
npm run lint
npm run test:integration
npm run test:live

# Compile and watch (TypeScript only; re-run build:backend after C# changes)
npm run watch

# Build just the backend
npm run build:backend

# Package the extension (TypeScript + backend)
npm run package

# Build a VSIX for local install / Marketplace submission
npm run package:vsix

npm run test:integration launches the extension in an Extension Development Host and verifies activation plus command registration.

npm run test:live does the same, but seeds a temporary VS Code profile from your current settings so it can exercise a real LM Studio connection with the configured lmstudio-copilot-expansion.apiKey.

License

GPL-3.0

LM Studio Copilot Expansion

My Best Tools