CLLMs for Copilot Chat

中文

Getting Started

Prerequisites

• VS Code 1.116 or later. This extension uses VS Code's Language Model Chat Provider integration. If you encounter compatibility issues with a specific VS Code/Copilot build, please report an issue.
• GitHub Copilot subscription (Free / Pro / Enterprise — the free tier works)
• An API key for at least one provider:
  • Qwen: a DashScope (Alibaba Cloud Model Studio) key from Model Studio, or any compatible token when using a custom cllms.baseUrl.
  • DeepSeek: an API key from the DeepSeek platform.
  • z.ai (GLM): a key from the z.ai API keys page.
  • MiniMax: a key from the MiniMax platform.
  • Xiaomi MiMo: a pay-as-you-go (sk-...) key from the Xiaomi MiMo open platform.
  • Moonshot (Kimi): a key from the Moonshot console (international; platform.moonshot.cn for mainland China).
  • Tencent Hunyuan (混元): a key from the Tencent Cloud Hunyuan console.

Installation

Install from the registry used by your editor:

1. Microsoft VS Code — install from VS Code Marketplace.

Usage

1. Run CLLMs: Set API Key from the Command Palette (Cmd/Ctrl+Shift+P) and pick a provider
2. Paste that provider's API key or compatible token (Qwen DashScope keys usually start with sk-)
3. Open Copilot Chat, click the model picker, pick a model
4. That's it — chat away

Remote Development

CLLMs works with Remote - SSH, Dev Containers, and WSL. The extension runs on the local client side, so:

• Install the extension locally (no need to install it on the remote machine).
• API keys stored in the local SecretStorage are used automatically — no reconfiguration needed.
• Settings (cllms.*) are read from your local settings.json.

Models

Seven providers ship out of the box. Each model has its own API key and endpoint, so you can use Qwen, DeepSeek, z.ai (GLM), MiniMax, Xiaomi MiMo, Moonshot Kimi, Tencent Hunyuan, or any combination at the same time from the Copilot model picker.

Qwen (DashScope)

DeepSeek

z.ai (Zhipu GLM)

MiniMax

Xiaomi MiMo

Moonshot (Kimi)

Tencent Hunyuan (混元)

Model IDs are the official provider names and are fully configurable via cllms.modelIdOverrides / cllms.zai.modelIdOverrides / cllms.minimax.modelIdOverrides / cllms.xiaomi.modelIdOverrides / cllms.moonshot.modelIdOverrides / cllms.hunyuan.modelIdOverrides / cllms.deepseek.modelIdOverrides for third-party / self-hosted endpoints.

Adding a new model

Want to add your own model? See Adding a new model for a step-by-step guide.

Testing Status

💡 Help wanted! International endpoints share the same API surface as their domestic counterparts, so they should work out of the box — but they haven't been verified yet. If you have an international API key, please give it a try and report your results. If you'd like to contribute test tokens, reach out via GitHub Issues. Every bit of testing helps make these providers more reliable for everyone.

Settings

Thinking Effort is configured from Copilot Chat's model picker for each thinking-capable model.

See Advanced settings for a focused guide to the cost and cache controls above.

For text-only models, the vision proxy resolves image attachments to text before the main request. All CLLMs models accept pasted images — text-only models auto-route them through the vision proxy, while native-vision models (e.g. Qwen3-VL-Plus, GLM-5V-Turbo) receive images directly. If no explicit vision model is configured, the extension falls back to any available Copilot model (e.g. GPT-4o) as a vision describer. Identical image descriptions are cached for the current VS Code session by image bytes, prompt, and describer identity, so retries or re-attaching the same image avoid another description call. API-endpoint vision calls also use cllms.maxRetries for transient 429 / 5xx / network failures and cllms.visionProxy.timeoutMs for each attempt.

For audio attachments, the audio proxy can transcribe audio into text before the main request for models that do not natively process audio parts. Configure it from CLLMs: Configure Audio Proxy (panel UI, similar to Vision Proxy). API-endpoint audio calls use the same retry policy (cllms.maxRetries) and cllms.audioProxy.timeoutMs per attempt.

Recent audio hardening updates:

• Quick Setup now validates and normalizes audio proxy endpoint config before saving (same behavior as the panel flow).
• Custom audio proxy headers cannot override protected request headers (authorization, content-type).
• Responses-audio MIME handling now supports audio/webm and audio/m4a/audio/mp4 explicitly; unsupported MIME types return a clear typed error.

Utility cost control

CLLMs can run lightweight, one-shot Copilot helper requests more cheaply than your real agent turns. Two independent paths exist, with different decision-makers.

Who decides a request is "utility"? CLLMs classifies every incoming request automatically from its system-prompt signature and tool set — this is not user-configurable. The eight utility-tier kinds are chat-title, git-commit-message, git-branch-name, inline-progress-message, rename-suggestions, prompt-categorizer, settings-resolver, and todo-tracker. Everything else — your real agent turns (main-agent), terminal follow-ups, generic background work, and any unrecognized request — is agent-tier and is never capped or downgraded, so a real turn is never throttled by mistake.

Path A — CLLMs downgrade (cllms.utility.*). Once a request reaches a CLLMs model, a downgrade fires only when the request is utility-tier and you opted in:

• cllms.utility.maxOutputTokens caps output tokens for utility requests (combined with cllms.maxTokens by taking the smaller value).
• cllms.utility.modelIdByProvider swaps the API model to a cheaper one on the same provider/key (e.g. { "qwen": "qwen-flash" }).

Both are off by default — with no config, nothing changes. The model picker still shows your selected model (this is a silent, server-side swap), and each downgrade/cap is logged when cllms.debugMode is metadata or verbose. Session cost uses the override model's built-in pricing when CLLMs recognizes that model; otherwise the request is shown as unbilled.

Path B — VS Code native (chat.utilityModel / chat.utilitySmallModel). Here VS Code / Copilot itself decides which requests are utility and routes them — before they reach any provider. CLLMs does not participate in that decision; it just serves whichever model the host picked. Run CLLMs: Configure Utility Model to set these native settings.

Prefer Path B when your VS Code supports it (routing is explicit and the cost estimate stays accurate); use Path A to fine-tune per provider or as a fallback.

Commands

Run these from the Command Palette (Cmd/Ctrl+Shift+P):

Note on session cost — Show Session Cost is an approximation for the current session only. It is estimated from the usage each provider streams back (cached input is billed at the cache-hit tier), is not a substitute for your provider's official billing, and resets when the display currency changes. The detail also reports the session average context-cache hit rate and, when utility requests incurred billed cost, the utility vs agent split. Requests for models without configured pricing are counted separately as "excluded for lack of pricing" rather than folded into the total, so the figure never looks more complete than it is.

Example settings.json override for compatible API proxies:

{
  "cllms.modelIdOverrides": {
    "qwen3-coder-plus": "your-coder-model-id",
    "qwen-plus": "your-plus-model-id",
    "qwen3-max": "your-max-model-id",
    "qwen3-vl-plus": "your-vl-model-id"
  }
}

Thanks

CLLMs began as a Qwen-focused adaptation of Vizards/deepseek-v4-for-copilot by Vizards, which pioneered the approach of plugging a BYOK model into the Copilot Chat picker via the native LanguageModelChatProvider API, and has since grown into a multi-provider extension for Qwen, DeepSeek, z.ai (GLM), MiniMax, Xiaomi MiMo, Moonshot Kimi, and Tencent Hunyuan. Huge thanks to the original author — the provider pipeline, vision proxy, thinking-mode handling, and diagnostics here are deeply inspired by the generous foundation that Vizards created and shared with the community.

License

MIT — see NOTICE for attribution to the upstream project.