Sockless LLM Router — Copilot Chat Provider
Use models served by your local Sockless LLM Router as Copilot Chat models in VS Code. Point the extension at the router's API endpoint and its models appear in the Copilot model picker automatically — no manual per-model configuration.
Sockless LLM Router is a local gateway that manages and launches model server presets (llama.cpp, etc.) and exposes them behind OpenAI- and Anthropic-compatible APIs. See the project page and GitHub repo for setup and preset configuration.
Screenshots
Presets configured in the router's admin UI (Presets) show up automatically in the Copilot Chat model picker, with context size and tool/vision support carried over:
Reasoning effort is set globally from the Command Palette or status bar, and only applied to presets that support it:
Setup
- Install the extension and run Sockless LLM Router: Configure Connection from the Command Palette (or open the Copilot model picker — it will prompt you the first time).
- Enter the router's API endpoint URL (the API port, e.g.
http://localhost:5054 — not the admin UI's 5053).
- Optionally enter an API key, only needed if the router's gateway has "Require API Key" enabled. The key is stored in VS Code's secret storage, never in settings.
- Open the Copilot Chat model picker — every server preset configured in the router shows up as a model, with the model picker reflecting each one's context size and whether it supports tools/vision.
To change the endpoint or key later, run Sockless LLM Router: Configure Connection again.
How model discovery works
On connect, the extension calls the router's GET /v1/models/capabilities endpoint, which reports — per preset — context length, max output tokens, and whether tool calling / image input are supported. That's what lets VS Code populate the model picker without asking you to describe each model by hand.
Reasoning effort
VS Code's model picker has no per-model "submenu" UI for a BYOK provider the way it does for Copilot's own GPT-5/o-series models — every entry provideLanguageModelChatInformation returns is just a flat, separate row in the picker. Rather than clutter the picker with a near-duplicate row per effort level, reasoning effort is a single global setting, shown and changed from the Reasoning: … status bar item (bottom right) — click it, or run Sockless LLM Router: Set Reasoning Effort from the Command Palette, to pick Auto/Low/Medium/High/XHigh.
The chosen level is sent as reasoning_effort on /v1/chat/completions, but only for a request to a preset the router reports as supporting it (supports_reasoning_effort) — it's silently omitted for any other model, so the one global setting never affects a model that doesn't understand it. A preset that does support it shows reasoning in its picker detail text as a hint.
This only applies to the openai protocol — the router doesn't yet translate a reasoning-effort level onto its claude-protocol /v1/messages path, so the setting has no effect while Sockless LLM Router: Protocol is claude.
Protocol
The router speaks two chat protocols side by side, and this extension can use either one — set Sockless LLM Router: Protocol (socklessLlmRouter.protocol) to:
openai (default) — POST /v1/chat/completions, the router's OpenAI-compatible endpoint.claude — POST /v1/messages, the router's Anthropic-compatible endpoint.
Both protocols expose the same presets with the same tool-calling and image-input support, so this is mainly for comparing behavior between the two or matching another Claude-protocol client talking to the same router.
Requirements
- A running Sockless LLM Router instance with at least one server preset configured.
- VS Code 1.104 or newer (for the Language Model Chat Provider API).
Links
Development
npm install
npm run compile
Press F5 in VS Code to launch an Extension Development Host with the provider loaded.
