LM-CODE

LM-CODE lets you use one or more LM Studio servers as models inside GitHub Copilot Chat in VS Code.

What You Get

• Connect multiple LM Studio servers.
• Full tool-calling (agent mode) support, including models that emit native tool-call markup instead of OpenAI-format tool calls. Recognized dialects: Hermes/Qwen <tool_call> JSON, Qwen3-Coder XML (<function=/<parameter=), Gemma ```tool_code Python calls, DeepSeek R1/V3/V3.1 special tokens, Mistral [TOOL_CALLS], Llama <|python_tag|>, Kimi K2 section tokens, GLM <arg_key>/<arg_value>, MiniMax/Claude-style <invoke> XML, Granite <|tool_call|>, LM Studio's [TOOL_REQUEST] bridge, bare-JSON calls, and Python dict literals — so Llama, Qwen, Mistral, Grok, DeepSeek, Kimi, GLM, Gemma, Granite, MiniMax and similar local models work out of the box.
• Capability-aware model discovery via LM Studio's REST API: real context lengths, native-vs-bridged tool support shown per model, and image input for vision models.
• Reasoning models (Qwen 3, DeepSeek R1, etc.) handled correctly: thinking output is filtered out of chat responses.
• Robust streaming: mid-stream server errors are reported clearly instead of producing silent empty responses, and transient model-load failures are retried automatically.
• Generous timeouts for just-in-time model loading (large models can take minutes to load before the first token).
• LM Link aware: models mirrored across your machines are de-duplicated in the picker, with per-server overrides and automatic failover to another server that has the model.
• Test each server connection directly from the extension UI.
• Auto-refresh available models.
• Hide models you do not want in your picker.
• Import and export full configuration JSON.
• View all servers and models in a dedicated sidebar.

Prerequisites

Before installing LM-CODE, make sure you have:

1. VS Code 1.95.0 or newer.
2. GitHub Copilot and Copilot Chat enabled in VS Code.
3. LM Studio installed and running.
4. At least one model loaded in LM Studio.
5. LM Studio API endpoint enabled (default is usually http://localhost:1234).

Quick Start (Recommended)

1. Install LM Studio and load a model.
2. Verify LM Studio API is reachable at http://localhost:1234/v1/models.
3. Install LM-CODE.
4. Open Command Palette and run LM-CODE: Open Settings Panel.
5. Add your server URL and click Test connection.
6. Open Copilot Chat and select an LM Studio model.

Installation

You can install LM-CODE in two ways.

Option A: Install from a VSIX release file

1. Download the latest .vsix from this repository's Releases page.
2. Install using VS Code:

code --install-extension lm-code-0.9.0.vsix

Option B: Build and install locally

npm install
npm run compile
npm run package
code --install-extension lm-code-0.9.0.vsix

Configure LM-CODE

Use the Settings Panel (easiest)

1. Run LM-CODE: Open Settings Panel.
2. Click Add Server.
3. Fill in:
   • Display name: friendly name (example: Local LM Studio)
   • Base URL: LM Studio endpoint (example: http://localhost:1234)
   • API key: optional (leave empty for local default setup)
   • Timeout: default 300000 (5 min — large models can take a while to load on first request)
4. Click Test connection.
5. Click Refresh Now to pull models.
6. Uncheck models you want hidden from Copilot picker.

Configure in VS Code settings.json

LM-CODE settings are under lmstudioCopilot.*.

Example:

{
  "lmstudioCopilot.refreshIntervalSec": 60,
  "lmstudioCopilot.showStatusBar": true,
  "lmstudioCopilot.servers": [
    {
      "id": "local",
      "name": "Local LM Studio",
      "baseUrl": "http://localhost:1234",
      "apiKey": "",
      "enabled": true,
      "timeoutMs": 300000,
      "headers": {},
      "refreshIntervalSec": 0,
      "hiddenModels": []
    }
  ]
}

LM Link setups

LM Link makes every machine on the link report the whole network's models. Two recommended ways to use it with LM-CODE:

One hub server (simplest). Add a single server — your local machine that is on the LM Link. Its listing already includes every model on the network, LM Studio routes each request to the right device, and the picker shows no duplicates.

Multiple servers. Add each machine as its own server. Because every machine lists the whole network, LM-CODE de-duplicates by model identity: each model is provided by the first enabled server that lists it (drag your preferred server to the top of the lmstudioCopilot.servers array to prioritize it). The sidebar shows how many duplicate models were hidden per server — click that row (or use the per-server "Show duplicate models" toggle in the settings panel) to keep a server's copies visible, labeled as duplicates. Turn the global behavior off with:

{
  "lmstudioCopilot.dedupeAcrossServers": false
}

Failover. If the providing server can't serve a model (LM Link drop, model not found, load failure) and nothing has been streamed yet, LM-CODE automatically retries the request on every other configured server that lists the same model.

Note: LM Studio's REST API does not expose which LM Link device a model lives on, so de-duplication is by model identity (the same model name on two servers is treated as one model). For the same reason, hiding a model while de-dupe is on hides that model identity across all servers — otherwise it would just reappear from a mirror.

Big context windows (256K → 1M)

Context size is fixed when LM Studio loads the model — no API request can change it. LM-CODE advertises to Copilot whatever the model is actually loaded with.

1. Load the model with a bigger window. LM Studio's GUI context slider stops at the model's native maximum, but the CLI does not:

   lms load "qwen/qwen3.6-35b-a3b" -c 1048576

   Watch memory: KV cache grows linearly with context and can dwarf the model weights. Use --estimate-only first to preview the cost.

2. LM-CODE picks it up automatically — discovery reads the actually-loaded context from LM Studio's API on the next refresh and budgets Copilot prompts accordingly.

3. Or pin it manually with a per-server override (settings panel → server → "Context overrides", or in settings.json):

   {
     "contextOverrides": { "qwen/qwen3.6-35b-a3b": 1048576 }
   }

Quality note: running far past the native window without RoPE/YaRN scaling weakens long-range recall. Models with a trained 1M variant (or YaRN-tuned GGUFs) behave much better at extreme contexts than a plain metadata override.

Commands

• LM-CODE: Open Settings Panel
• LM-CODE: Refresh Models Now
• LM-CODE: Add Server
• LM-CODE: Export Configuration
• LM-CODE: Import Configuration

Troubleshooting

No models appear

1. Confirm LM Studio is running.
2. Confirm a model is loaded in LM Studio.
3. Open http://localhost:1234/v1/models in browser.
4. Click LM-CODE: Refresh Models Now.

Test connection fails

1. Verify Base URL is correct and does not end with extra path segments.
2. If using remote server, confirm firewall/network access.
3. If your server requires auth, set API key or custom headers.
4. Increase timeout from 300000 if your host is slow.

Copilot Chat does not show LM Studio models

1. Ensure GitHub Copilot Chat is installed and enabled.
2. Reload VS Code window.
3. Re-open LM-CODE panel and refresh models.
4. Ensure server is enabled in LM-CODE settings.

Tool calls fail or look wrong for a particular model

Hover the model in the picker: "native tools" means the model's chat template handles tool calls itself; "bridged tools" means LM Studio injects a prompt-based tool format, which is more fragile on small or heavily quantized models. Check each model's reported capabilities at http://localhost:1234/api/v0/models.

The response is empty or cuts off

• Check the LM-CODE output channel (View > Output > LM-CODE) — mid-stream server errors (for example "Model unloaded.") are logged there and surfaced in chat.
• Reasoning models think before answering; the thinking is hidden. If a model spends its entire output budget thinking, LM-CODE tells you in the response.

"No utility model is configured for 'copilot-utility-small'..." (VS Code 1.128+)

VS Code needs a small "utility" model for background chores (chat titles, query rewriting) and defaults it to "none" when your main agent model is a bring-your-own model like LM-CODE's. Fix it in settings:

{
  "chat.byokUtilityModelDefault": "mainAgent"
}

"mainAgent" keeps utility calls on your selected local model (fully local). Use "copilot" instead if you are signed in to GitHub Copilot and prefer its hosted utility models for those background tasks.

First request to a model is slow or times out

LM Studio loads models just-in-time; large models can take minutes before the first token. LM-CODE waits up to the per-server timeout (default 5 minutes) and automatically retries once when LM Studio reports a transient load failure.

Security Notes

• Keep API keys local to your machine.
• Do not commit private keys or secrets to source control.
• Exported config files may contain API keys; treat them as sensitive.

Developer Notes

Build commands:

npm install
npm run compile
npm run lint
npm run package

License

MIT