"cliproxy.customModelParameters": [
  {
    "pattern": "openai::gpt-5*",
    "maxInputTokens": "1m",
    "maxOutputTokens": "64k",
    "imageInput": true,
    "toolCalling": true
  },
  {
    "pattern": "openai::gpt-4*",
    "maxInputTokens": "128k",
    "maxOutputTokens": "16k"
  },
  {
    "pattern": "claude::claude-3.7-*",
    "maxInputTokens": "200k",
    "maxOutputTokens": "64k"
  }
]

Introduction

CLIProxy for Copilot is a VS Code extension that registers CLIProxyAPI as a custom Language Model Chat Provider for VS Code.

With this extension, you can use various large language models (such as Gemini 2.5 Pro, GPT-5, Claude 3.7 Sonnet, etc.) proxied by CLIProxyAPI directly within VS Code Copilot Chat. It allows you to leverage your existing OAuth subscriptions or free quotas without needing expensive API keys.

Key Features

• Automatic Model Discovery: Automatically fetches the list of available models from your CLIProxyAPI server.
• Native Integration: Switch models directly in the Copilot Chat model picker for a seamless experience.
• Streaming Responses: Supports SSE streaming for fast and responsive interactions.
• Tool Calling Support: Fully compatible with Copilot's tool calling capabilities.
• Easy Configuration: Quickly configure server settings via commands or the settings UI.
• Status Bar Shortcut: Adds a CLIProxy icon button to the right side of the VS Code status bar for quickly opening the settings panel.
• Built-in SQL.js Runtime: Usage statistics persistence ships with built-in sql.js runtime assets under resources/vendor/sql.js/, so the published extension no longer depends on runtime files under node_modules/sql.js.

Quick Start

1. Install the Extension: Install this extension in your VS Code.
2. Configure Server:
   • Run the command CLIProxy: Configure Server URL and API Key.
   • Enter your CLIProxyAPI server URL (default: http://localhost:8317).
   • Enter the API Key if your server requires authentication.
3. Use Models:

• Open the Copilot Chat window.
• Click the model picker and select a model provided by CLIProxy.
• Start chatting!

Extension Settings

This extension contributes the following settings:

• cliproxy.serverUrl: The URL of your CLIProxyAPI server.
• cliproxy.apiKey: The API Key required to access the server (optional).
• cliproxy.requestTimeout: Request timeout in seconds.
• cliproxy.logUsageStats: Whether to log the final response usage JSON to the CLIProxy output channel when usage stats are present.
• cliproxy.useGeminiProtocolForGoogleModels: For google or antigravity models whose ID contains gemini, use Google Gemini AI Studio /v1beta/models/{model}:generateContent for chat requests only; model discovery still uses the existing OpenAI-compatible /v1/models endpoint.
• cliproxy.replaceManageTodoListForGemini: When enabled, Gemini user chat agent requests that include manage_todo_list replace it with cfc_todo_init and cfc_todo_set_status before sending to the model, then internally forward model calls back to manage_todo_list; hidden metadata in tool call IDs restores history for Gemini and is stripped for non-Gemini requests.
• cliproxy.forceCurrentModelForSubAgent: If enabled, the model and agentName parameters will be removed when the AI calls runSubagent, forcing the sub-agent to use the current model.
• cliproxy.disableTaskComplete: If enabled, the task_complete tool function will be disabled, removed from request tools, and the completion instruction text in the first message will also be removed.
• cliproxy.logLevel: Output panel log level. Available values: debug / info / warn / error / metric. Only logs at or above the selected level are shown.
• cliproxy.customModels: Custom model list. Each entry can define id, optional display name name, request url, and optional dedicated key. The url can be a base endpoint that automatically appends /v1/chat/completions, or a full /v1/chat/completions URL that is used as-is. If the dedicated key is left empty, requests for that custom model will not fall back to the global cliproxy.apiKey.
• chat.utilityModel: Copilot utility model for helper tasks such as Git commit message generation. Must use cliproxyapi/<provider>::<modelId> format.
• chat.utilitySmallModel: Copilot lightweight utility model for small tasks such as chat title generation. Must use cliproxyapi/<provider>::<modelId> format.
• cliproxy.customModelParameters: Ordered custom model parameter rules array using model IDs or wildcard patterns in <provider>::<modelId> form. A single * matches all providers and models. It supports maxInputTokens, maxOutputTokens, imageInput, and toolCalling; omitted values default to 1m input, 64k output, image input supported, and tool calling supported.
• cliproxy.toolFilters: Tool function filter rules. During request initialization, the extension collects VS Code options.tools declarations, saves them to tool-functions.json in global extension storage, and uses them for autocomplete in the Settings → Filter Settings tab. When the VS Code version changes, the stored list is cleared and recollected from future requests. Tool schemas are cached as independent snapshots to avoid in-place filtering from polluting the cache and causing frequent writes. Each rule supports one or more model wildcards, removing one or more tools, or rewriting tool / parameter descriptions. Tool and parameter targets support partial name or description matching. The tool list file uses a lock file plus atomic temp-file replacement to avoid multi-process write conflicts. Manual clearing writes a resetAt marker so other VS Code processes discard stale in-memory caches and recollect tools.

Example:

"cliproxy.customModelParameters": [
  {
    "pattern": "openai::gpt-5*",
    "maxInputTokens": "1m",
    "maxOutputTokens": "64k",
    "imageInput": true,
    "toolCalling": true
  },
  {
    "pattern": "openai::gpt-4*",
    "maxInputTokens": "128k",
    "maxOutputTokens": "16k"
  },
  {
    "pattern": "claude::claude-3.7-*",
    "maxInputTokens": "200k",
    "maxOutputTokens": "64k"
  }
]

pattern uses the provider-qualified <provider>::<modelId> form directly, such as openai::gpt-5*. You can still use * for bulk matching, such as openai::gpt-4*.

For Copilot-specific helper models, the settings page also allows configuring chat.utilityModel and chat.utilitySmallModel. Their values must use the cliproxyapi/<provider>::<modelId> form, such as cliproxyapi/google::gemini-3.1-flash-lite-preview. Both fields now support dropdown suggestions based on the currently available model list.

For custom models, requests use the custom entry's url and dedicated key. A base endpoint automatically appends /v1/chat/completions; a full /v1/chat/completions URL is used as-is. If the dedicated key is empty, the request is sent without an Authorization header and does not fall back to the global cliproxy.apiKey.

For tool filters, the Settings → Filter Settings tab shows affected models for each wildcard rule and offers autocomplete from the saved tool / parameter list. If a rule both removes tools and rewrites descriptions, removal is applied first, then rewrites are applied to remaining tools.

When cliproxy.useGeminiProtocolForGoogleModels is enabled, Gemini text-part thoughtSignature values are preserved in chat history as compact [gemini_sign](https://github.com/easyatm/vscode-extension-CLIProxyCopilot/blob/HEAD/resources/geminiSignatureHint.md "base64-signature") markers and restored in later Gemini requests. Non-Gemini requests strip these markers automatically. See resources/geminiSignatureHint.md for details.

When cliproxy.replaceManageTodoListForGemini is enabled, only UserChat agent requests for model IDs containing gemini are affected. If the request declares manage_todo_list, it is replaced with cfc_todo_init / cfc_todo_set_status, and tool calls are forwarded back to VS Code's manage_todo_list with a complete todoList payload.

For logging, most current output panel messages are recorded at debug level by default, while usage statistics are recorded at metric level. Since metric is higher than error, setting the log level to metric will only show these highest-priority metric lines. The usage metric line also includes requestBodyBytes, representing the byte size of the outgoing chat request body.

Troubleshooting

• Q: Why don't I see CLIProxyAPI models in the list?
  • A: Ensure your CLIProxyAPI server is running and the URL is correctly configured. You can run the CLIProxy: Refresh Available Models command to trigger a manual refresh.
• Q: Does it support image input?
  • A: The extension automatically infers multimodal capabilities based on the model name. Models that support vision can handle image inputs directly.

License

MIT