Copilot Models

Unlock third-party large language model extensions for GitHub Copilot.

Seamlessly integrate DeepSeek, Zhipu AI, and Qwen LLMs.

One-click switching and native panel compatibility.

Features

• Multi-Model Support: DeepSeek V4, Zhipu AI GLM-5, Qwen 3 series
• Model Routing: Automatic failover and latency-based routing
• Tool Calling: Full Copilot Chat tool calling support
• Thinking Mode: Model reasoning/thinking mode support
• Vision Proxy: Image description proxy for non-vision models via VS Code LM or custom API
• Circuit Breaker: Automatic failure protection with retry
• Secure Authentication: API keys stored in VS Code SecretStorage; sensitive values (keys, tokens, URLs) are automatically redacted from logs
• Log Debugging: 4-level logging with hot-reload
• Lightweight: OpenAI SDK replaced with native SSE client code
• Token Plan: Unified prepaid billing for Qwen, DeepSeek, and GLM token packages via a single endpoint
• Utility Model: Optionally run AI commit-message generation on one of these models, through VS Code's utility-model settings

Documentation

Quick Start

1. Install Extension

Install from the VS Code Extension Marketplace.

2. Configure API Key

Press Ctrl+Shift+P (macOS: Cmd+Shift+P), run Copilot Models: Set API Key, select a provider and enter your key.

Note: API keys are stored in VS Code SecretStorage, not as plain settings.

3. (Optional) Configure Token Plan

If you use prepaid token packages (e.g., Alibaba DashScope plan), run Copilot Models: Set Token Plan to configure plan access:

1. Press Ctrl+Shift+P (macOS: Cmd+Shift+P), run Copilot Models: Set Token Plan
2. Select a built-in provider preset or enter a custom URL
   • The Qwen preset is preconfigured with the endpoint URL and 6 supported models
3. Enter the plan API token
4. Select the models covered by this plan

The plan token is saved in VS Code SecretStorage.

• Qwen Token Plan — https://token-plan.cn-beijing.maas.aliyuncs.com/compatible-mode/v1

The Qwen Token Plan preset covers Qwen, DeepSeek, and GLM models in a single plan. For other providers, choose "Custom URL" and enter the plan API endpoint.

Run Copilot Models: Clear Token Plan to remove a configured plan.

4. (Optional) Configure Vision Model

If you want to use image attachments with models that don't natively support image input (e.g., GLM-5.3, GLM-5.2), configure a vision proxy to automatically convert images to text descriptions:

1. Press Ctrl+Shift+P (macOS: Cmd+Shift+P), run Copilot Models: Set Vision Model
2. Select a vision-capable model, or choose "Custom API Endpoint"
3. For custom API endpoint, enter the URL, model ID, and API key (leave the key empty for unauthenticated endpoints)

The vision proxy describes images before sending them to the chat model. For custom API endpoints, an OpenAI-compatible endpoint is required — enter either the base URL (https://host/v1) or the full /chat/completions URL, both are accepted. The API key is stored in VS Code SecretStorage.

Note: The proxy only applies to models that cannot accept image input natively. Models with image support receive the original images and are never routed through the proxy.

Only the most recent message is described. Images from earlier turns — and any image whose description fails or is empty — are replaced with [Image: description unavailable], so no image part reaches a model that cannot accept one.

Run Copilot Models: Clear Vision Model to remove the configuration.

5. Start Using

1. Open GitHub Copilot Chat panel
2. Click on the model selector
3. Select the model to use
4. Start chatting

Supported Models

The Thinking Effort selector in the model picker maps onto what each provider documents. None turns thinking off everywhere: Qwen through enable_thinking: false — its API defaults thinking to on — and Zhipu AI and DeepSeek through their thinking toggle. All three document an on/off switch and no effort level, so low, high and max enable thinking at the provider's own default effort.

Qwen (Alibaba Cloud)

DeepSeek

Note: deepseek-v4-flash has been retired and deepseek-v4-pro is being retired — requests to either ID are served by DeepSeek-V4.1-Flash.

Zhipu AI (BigModel)

Tip: Models marked with ❌ for Image Input can still handle images through the Vision Proxy feature (see Quick Start step 4).

Token Plan Coverage

The built-in Qwen Token Plan preset supports the following models through a single unified endpoint:

Models not listed (e.g. GLM-5-Turbo, kimi-k2.7-code) are still available via direct provider API access — they are simply not covered by this Token Plan preset.

Note: The IDs above must match the model IDs exposed by the extension (see the Supported Models tables). deepseek-v4-pro and deepseek-v4-flash are accepted by the DeepSeek API itself, but the extension only exposes them as deepseek-flash.

Configuration Options

Available in VS Code settings (search copilot-models):

Provider Settings

Global Settings

On editTools: left empty (the default), the editor tries several edit tools and picks one itself. Accepted values are find-replace, multi-find-replace, apply-patch and code-rewrite; fill them in only when you know which editing tool suits the models, because naming the wrong one makes their edits worse. The setting applies to every model.

Vision Proxy Settings

Note: visionProxy.maxTokens applies only to the vision description request sent to visionProxy.apiUrl — it does not affect normal chat requests. Chat requests have no separate max_tokens setting: each model automatically uses its own maxOutputTokens as the API's max_tokens parameter. See the "Output" column in the Supported Models tables above.

Token Usage

Every completed request records its token usage locally — including requests that use a directly configured API key, not only token plan traffic.

• The status bar shows today's tokens and request count, for example 12.3K tok · 18 req. Click it to open the report. The item stays hidden until the first request is recorded; set copilot-models.showStatusBar to false to hide it permanently.
• Run Copilot Models: Show Token Usage for a breakdown by plan and by model.
• Run Copilot Models: Clear Token Usage to drop the recorded history.

Note: Only the most recent 1000 requests are kept, so the "all time" figures are a rolling window rather than a lifetime total. Usage is stored in VS Code global state and never leaves your machine.

Account Balance

The usage report also shows your DeepSeek account balance, queried from the official GET /user/balance endpoint. It refreshes each time you run the command, and is shown even before any request has been recorded.

Balance:
  deepseek: ¥110.00 (granted ¥10.00 · topped up ¥100.00)

Note: DeepSeek is the only supported provider with a documented balance API. Zhipu AI, Qwen/DashScope and the Qwen Token Plan endpoint expose none, so no figure is shown for requests served through them. Providers without a configured API key are omitted entirely; a configured provider whose lookup fails reads as unavailable — a balance problem never blocks the rest of the report. No request is sent when no API key is set.

AI Commit Messages

The sparkle button in the Source Control input box drafts a commit message. The GitHub Copilot Chat extension provides that button, and by default it runs on Copilot's own small utility model. It can run on a model from this extension instead.

Point it at one by setting chat.utilitySmallModel to the <vendor>/<model-id> value the dropdown stores. Every model from this extension is served by the Copilot Models router, whose vendor id is copilot-models-router:

"chat.utilitySmallModel": "copilot-models-router/deepseek-flash"

Note: The vendor is the one that registers the models, not the upstream service. deepseek/deepseek-flash looks right and is silently ignored — the setting is only read by name <vendor>/<id>, and no model is registered under the deepseek vendor. Picking from the dropdown avoids the guesswork.

chat.byokUtilityModelDefault decides what happens when the model selected in the chat picker is a BYOK model and neither override is set: copilot (the default) keeps Copilot's utility models, mainAgent reuses the selected BYOK model, and none disables utility models.

Notes:

• The model must be selectable, which means its API key or a covering token plan is configured. Without credentials a model is not offered in these settings, and a request it does receive fails with API key not configured.
• Prefer a fast, inexpensive model: these flows run often and the prompt is mostly the diff.
• Only the Copilot extension's utility aliases follow these settings. VS Code's own internal flows (chat titles, dictation cleanup, tool risk assessment) ask for the Copilot utility model directly and ignore them.
• To change what the message says rather than which model writes it, use github.copilot.chat.commitMessageGeneration.instructions.

Commands

Debugging

If you encounter issues, check the logs:

1. Press Ctrl+Shift+P, run Copilot Models: Show Log
2. Logs appear in the "Copilot Models" output panel
3. Set debugMode to verbose for detailed debug output
4. Set to minimal (default) for warnings and errors only

Log level changes take effect immediately without reloading the extension.

Each request's logs carry a structured prefix (req=<id> provider=<id> model=<id>), so you can grep for a single req=<id> to trace one request across routing, provider, and network layers. API keys, tokens, and URL query strings are automatically redacted from the logs — secrets never appear in the output panel.

Diagnosing images

Any request that carries an image writes one line describing what became of it, for example:

[WARN ] [Vision] req=0043ef provider=deepseek model=glm-5.3 \
  Image handling: found=1 inMessages=1 described=0 failed=1 unavailable=0 omitted=0

A request with no images writes no such line. When failed or unavailable is non-zero the line is logged as a warning, so it appears at the default minimal level; a request whose images were all handled is logged at info, which needs metadata or verbose to show up.

License

Apache-2.0