LM Studio (BYOM) for VS Code

A Bring-Your-Own-Model provider that registers your local LM Studio server with VS Code's language model API (vscode.lm). Once installed, every loaded LM Studio model shows up in the GitHub Copilot Chat model picker — and is available to any extension that calls vscode.lm.selectChatModels({ vendor: 'lmstudio' }).

Built against the language-model provider contribution point introduced in VS Code 1.104.

Requirements

• VS Code 1.104 or newer
• LM Studio running with the Local Server enabled (default http://localhost:1234)
• At least one chat-capable model loaded in LM Studio

Features

• Auto-discovery of loaded models via LM Studio's /v1/models endpoint
• Streaming chat completions over Server-Sent Events
• Tool / function calling for models that support it (OpenAI-style delta tool calls, plus inline <tool_call> XML format used by Qwen / Llama)
• Vision input for multimodal models that advertise the vision capability
• Strips <think>...</think> reasoning traces from output
• Optional API key for remote / authenticated LM Studio servers

Settings

Commands

• LM Studio: Refresh Available Models — re-fetches the model list
• LM Studio: Check Server Connection — pings /v1/models and reports status

Development

npm install
npm run compile      # one-shot bundle to dist/extension.js
npm run watch        # incremental rebuild
npm run package:vsix # produce a .vsix to install locally

Press F5 in VS Code to launch an Extension Development Host with the provider registered.

How it works

The extension registers a LanguageModelChatProvider under the vendor identifier lmstudio (declared in package.json under contributes.languageModelChatProviders):

vscode.lm.registerLanguageModelChatProvider('lmstudio', provider);

The provider implements three methods:

• provideLanguageModelChatInformation — returns the list of models that VS Code should show in the picker
• provideLanguageModelChatResponse — converts VS Code messages to OpenAI format, opens a streaming POST to /v1/chat/completions, and forwards text + tool calls back to VS Code as LanguageModelTextPart / LanguageModelToolCallPart
• provideTokenCount — character-based heuristic (LM Studio does not expose a tokenizer endpoint)

License

MIT