Custom LLM Router for VS Code

Use local and third-party models in GitHub Copilot Chat. Custom LLM Router registers Anthropic (Claude), Azure OpenAI, Ollama, LM Studio, vLLM, LocalAI, OpenRouter, OpenAI, Gemini, DeepSeek, Groq, Together, Mistral, GitHub Models, FreeLLMAPI, OmniRoute, or any OpenAI-compatible server as a native VS Code language model provider. You get streaming, tool calling (agent mode), image input and reasoning output.

Features

• Native in Copilot: the models you pick appear in the Copilot Chat model picker under Custom LLM Router.
• Three wire protocols:
  • OpenAI-compatible: most servers.
  • Ollama native (/api/chat): sends the context length (num_ctx) so long chats aren't cut short, plus keep-alive and native thinking.
  • Anthropic Messages API: uses the official SDK, with model limits from the Models API, summarized adaptive thinking, and refusal fallbacks on supported models.
• Fallback routes: combine several models into one picker entry. If the first is down, rate-limited or errors before answering, the next one takes over (for example, a local model first and a cloud model as backup).
• Health-checked: each model gets a quick chat check. With the full check it is also asked to call a tool, and the result is shown as calls tools, accepts tools but ignores them, or no tool support. Broken models never reach the picker.
• Accurate limits: real context windows are read from Ollama (/api/show) and LM Studio. You can override the name, context window, max output, tool support and image input for any model, and guessed values are marked ~.
• Cost-aware:
  • Before checking many models on remote (possibly paid) providers, it asks first and offers to check only your selected models. A basic mode halves the number of requests.
  • Bulk checks are rate-limited per provider.
  • Prices come from OpenRouter or Anthropic's list prices and can be overridden per model. Spend is estimated from reported usage.
• Real performance stats: time to first token and tokens per second, measured from your actual chats. They stay on your machine.
• Calibrated token counts: learned from the usage numbers the server reports, so Copilot trims context at the right point.
• Ollama extras: pull models from the dashboard or sidebar, and keep models loaded with a per-provider keep-alive.
• Secure keys: API keys and credential-like header values live in VS Code's encrypted secret storage, never in settings.json. Keys are stored per machine, so when a provider rejects the credentials on a new machine you're asked to enter the key there.
• Sidebar and dashboard: an Activity Bar view with each provider's health and per-model checkboxes, plus a dashboard for searching and filtering models, bulk actions, routes and settings. It handles thousands of models without slowing down.

Getting started

1. Install the .vsix:

   code --install-extension vscode-custom-llm-router-2.2.0.vsix --force
   

2. Open the LLM Router view in the Activity Bar and click Add Provider.
3. Pick a preset (or Custom), enter the API key if needed, then click Test connection and Add provider.
4. Click Verify untested to health-check the discovered models.
5. Tick the models you want, or create a Route, then select it in the Copilot Chat model picker.

Base URLs

The /v1 suffix is optional. The editor shows the exact chat URL it will call.

Under Advanced you can set:

• the API type and how the key is sent
• overrides for the chat and models URLs
• extra headers (credential-like values are stored securely)
• a timeout
• the Ollama context length and keep-alive
• model IDs for servers that don't list their models

Ollama context: with the default Ollama native API, models are loaded with the provider's context length (default 32k, capped at the model's maximum). Raise it if you have the memory; Copilot's limits follow it. If you switch an Ollama provider to the OpenAI-compatible API, Ollama uses OLLAMA_CONTEXT_LENGTH instead (often 4k–8k).

Settings

Commands

Right-click a provider in the sidebar to set its API key, test it, re-test its models, pull an Ollama model, edit, enable or disable, or delete it. Right-click a model to re-test it or edit its limits.

Troubleshooting

• "API key needed on this machine": API keys and secret headers are kept in each machine's secret storage and are not part of Settings Sync. Use Set API Key… (sidebar, dashboard or the notification).
• Anthropic refusals: if Claude declines a request you'll see the refusal category. On api.anthropic.com, Claude Fable 5.1, Opus 5.5, Opus 5 and Sonnet 5.5 use server-side refusal fallbacks (fallbacks: "default").
• Corporate proxy: requests use VS Code's http.proxy settings as long as http.fetchAdditionalSupport is enabled (the default). The extension warns you if a proxy is set but that option is off.
• A model is missing from the picker: it must be ticked (or part of a route) and have passed verification. Hover it in the sidebar to see why.
• Reasoning display: VS Code's native "thinking" UI is still a proposed API that Marketplace extensions can't use. Reasoning is shown as a quoted block instead, and that block is stripped from history so it isn't sent back to the model.

Upgrading

From 2.1: no action needed. Ollama providers now use Ollama's native API by default; set API type to OpenAI-compatible to keep the old behavior.

From 2.0: no action needed. Routes, overrides, and the Ollama and LM Studio limits are new. Tool support is re-classified the next time a model is checked.

From 1.x:

• Copilot selection is now opt-in. Your previously active models are carried over automatically.
• Profiles were removed. Use the filters, bulk Add shown to Copilot, or routes instead.
• Model IDs are now provider/model, so you may need to re-pick your model in the chat picker once.

Development

npm install
npm test                    # type-check + unit tests against the mock server + sidebar + dashboard (jsdom) tests
npm run test:integration    # launches a real VS Code with the extension and checks vscode.lm end to end
npm run bundle              # esbuild → dist/extension.js and media/dashboard.js (from src/webview/dashboard.ts)
npm run screenshots         # renders the dashboard in Dark/Light/High Contrast into test/visual/out (needs: npx playwright install chromium)
npm run l10n:export         # refresh l10n/bundle.l10n.json after changing user-facing strings
npx @vscode/vsce package --no-dependencies

Press F5 and choose Run Extension (with mock server) to try everything offline. mock_llm_server.py (port 31415) imitates:

• streaming, reasoning and <think> tags
• tool calls and usage reporting
• the Ollama and Anthropic APIs
• auth-protected paths
• slow and failing models