Why Matrix OAI Gateway?

Don't replace Copilot — power it up. This extension plugs directly into VS Code's language model provider API, so your custom models appear in the same model picker you already use every day. Agent mode, tool calling, Explore subagents, thinking mode — all preserved.

💡 Two-way gateway: it also exposes a local proxy (/v1/chat/completions, /v1/responses, /v1/messages), so external tools like Codex CLI, Continue, or curl can reach your models — or even VS Code's built-in Copilot models — through a single HTTP endpoint.

✨ Features

📸 Screenshots

🚀 Quick Start

1. Install the extension

Search "Matrix OAI Gateway" in the VS Code Extensions marketplace, or install from VSIX.

2. Add a provider & set API key

Open Command Palette (Ctrl+Shift+P):

Matrix OAI Gateway: Add Provider    → e.g. DeepSeek, https://api.deepseek.com/v1
Matrix OAI Gateway: Set API Key     → paste your key

3. Add a model (or use presets)

Matrix OAI Gateway: Add Preset Model  → pick from 15+ presets
Matrix OAI Gateway: Add Model         → custom model config

4. Start coding

Open Copilot Chat → click the model picker → select your model → done! 🎉

Model Visibility

VS Code 1.121 does not expose its native eye-icon visibility toggle to ordinary third-party language model providers. Matrix OAI manages picker visibility in its own configuration panel.

Open Command Palette (Ctrl+Shift+P):

Matrix OAI Gateway: Configuration

In OAI-Compatible Models, use the eye icon in the Action column to show or hide each model from the Copilot model picker.

Matrix OAI also supports default visibility in model config:

"matrixOaiCopilot.models": [
  {
    "id": "deepseek-v4-flash",
    "providerId": "deepseek",
    "showInModelPicker": true
  },
  {
    "id": "glm-5.1-anthr",
    "providerId": "zhipu-anthropic",
    "showInModelPicker": false
  }
]

Compatibility aliases are also accepted: hidden, visible, and isUserSelectable.

Hidden models stay configured for the local proxy, Codex config generation, Claude Code mappings, and vision proxy selection. They are only removed from the VS Code/Copilot model picker.

⚙️ Configuration

Providers & Models

Providers are reusable endpoints. Models reference providers by providerId, so one provider can host many models.

// settings.json
"matrixOaiCopilot.providers": [
  {
    "id": "deepseek",
    "name": "DeepSeek",
    "baseUrl": "https://api.deepseek.com/v1",
    "apiMode": "openai",
    "headers": {}
  },
  {
    "id": "ollama",
    "name": "Ollama Local",
    "baseUrl": "http://localhost:11434",
    "apiMode": "ollama",
    "headers": {}
  }
],
"matrixOaiCopilot.models": [
  {
    "id": "deepseek-v4-pro",
    "name": "DeepSeek V4 Pro",
    "providerId": "deepseek",
    "family": "deepseek",
    "maxInputTokens": 1000000,
    "supportsTools": true,
    "thinking": { "type": "enabled" },
    "thinkingFormat": "deepseek",
    "reasoning_effort": "max"
  },
  {
    "id": "qwen3.5:9b",
    "name": "Qwen3.5 9B",
    "providerId": "ollama",
    "family": "qwen",
    "maxInputTokens": 131072,
    "supportsTools": true,
    "supportsImages": true,
    "enable_thinking": true,
    "thinking_budget": 8192
  }
]

Key Settings

💰 Balance Checking

Query your API key balance directly from VS Code — no browser needed.

Supported providers:

How to use:

1. Command Palette → Matrix OAI Gateway: Check API Balance
2. Select a provider → enter API key (if not stored)
3. View balance in the status bar 🟢

Balance auto-refreshes after startup (30s delay) and every N minutes. The status bar shows a provider prefix (DS:, GLM:, OR:) so you know whose balance is displayed.

🧠 DeepSeek Thinking Mode

DeepSeek reasoning models need reasoning_content passed back in tool-calling turns. VS Code hides this field, so this extension:

• Stores reasoning_content in memory between turns
• Reattaches it automatically before upstream calls
• Falls back to empty reasoning_content when the original text can't be recovered

Model options:

🔌 Local Proxy

When enabled, the proxy listens at http://127.0.0.1:8080:

Routing logic: If model matches a configured OAI model → upstream provider. Otherwise → VS Code language model (e.g. Copilot).

Codex CLI Example

# ~/.codex/config.toml
[model_providers.deepseek]
name = "DeepSeek"
base_url = "http://127.0.0.1:8080/v1"
experimental_bearer_token = "codex-deepseek-local"
wire_api = "responses"

[profiles.deepseek-v4-pro]
model_provider = "deepseek"
model = "deepseek-v4-pro"

Or use the one-click command: Matrix OAI Gateway: Write Codex Config

Claude Code Example

Claude Code can use non-Claude models through the Anthropic API passthrough:

1. Add Anthropic provider + model (or use presets):

"matrixOaiCopilot.providers": [{ "id": "deepseek-anthropic", "baseUrl": "https://api.deepseek.com/anthropic", "apiMode": "anthropic" }],
"matrixOaiCopilot.models": [{ "id": "deepseek-v4-pro-anthr", "providerId": "deepseek-anthropic", "upstreamModelId": "deepseek-v4-pro" }]

2. Set model mapping:

"matrixOaiCopilot.proxy.anthropicModelMapping": {
  "claude-sonnet-4-20250514": "deepseek-v4-pro-anthr"
}

3. Run Matrix OAI Gateway: Write Claude Code Config → writes ~/.claude/settings.json
4. Claude Code will now use DeepSeek V4 Pro behind the scenes!

Model Mapping

Both Codex and Claude Code support model name mapping:

• proxy.codexModelMapping: { "o3": "deepseek-v4-pro" } — Codex sends o3, routes to DeepSeek
• proxy.anthropicModelMapping: { "claude-sonnet-4-20250514": "deepseek-v4-pro-anthr" } — Claude Code sends Claude name, routes to Anthropic API provider

🤖 Copilot Explore Subagents

GitHub Copilot Agent starts an internal Explore subagent for code search. This extension defaults it to your OAI model so it doesn't fall back to Copilot built-in models:

// Auto-configured by the extension
"chat.exploreAgent.defaultModel": "DeepSeek V4 Pro (matrix-oai-compatible)",
"chat.customAgentInSubagent.enabled": true

🎯 Preset Models (One-Click Setup)

📋 Commands

⏱️ Timeouts & Streaming

Slow reasoning models may need longer timeouts. Configure globally or per-model:

"matrixOaiCopilot.requestTimeoutSeconds": 300,  // global default

// Per-model override:
"matrixOaiCopilot.models": [
  {
    "id": "deepseek-reasoner",
    "requestTimeoutSeconds": 600,  // 10 min for thinking models
    "stream": true,
    "stream_options": { "include_usage": true }
  }
]

🔧 Compatibility

This extension supports:

• ✅ Any OpenAI-compatible /chat/completions endpoint
• ✅ Ollama native /api/chat (with tool-call support)
• ✅ VS Code language models via vscode.lm
• ✅ OpenAI-compatible proxy clients (curl, Continue, etc.)
• ✅ OpenAI Responses protocol (Codex CLI)
• ✅ Anthropic Messages passthrough (Claude Code)
• ✅ Model name mapping for both Codex and Claude Code

Tool calling, image input, reasoning options, and token usage depend on the upstream model.

📄 License

MIT

功能一览

Matrix OAI Gateway 可以把任意 OpenAI 兼容模型接入 VS Code / Copilot Chat，同时把 VS Code 里的语言模型反向暴露成本地 OpenAI / Anthropic 兼容接口。

核心特性：

• 🤖 Copilot Chat 无缝集成 — 模型直接出现在 Copilot 模型选择器里，支持 Agent 模式、工具调用、Vision 代理、思考模式
• 🔌 本地代理服务器 — 暴露 /v1/chat/completions、/v1/responses、/v1/messages 接口，支持 Codex CLI、Continue、curl 等外部工具
• 💰 余额查询 — 支持 DeepSeek、OpenAI、OpenRouter、智谱 GLM、SiliconFlow、Groq、Mistral 等 7+ 供应商，状态栏自动刷新
• 🧠 DeepSeek 思考模式 — 自动回放 reasoning_content，解决工具调用多轮对话的 invalid_request_error
• 🔀 Claude Code 第三方推理 — 通过 Anthropic API 透传让 Claude Code 使用 DeepSeek、GLM 等非 Claude 模型
• 🔀 Codex 模型映射 — Codex 指定别名模型名路由到任意配置模型
• 🔀 Claude Code 第三方推理 — 通过 Anthropic API 透传让 Claude Code 使用 DeepSeek、GLM 等非 Claude 模型
• 🔀 Codex 模型映射 — Codex 指定别名模型名路由到任意配置模型
• ⚡ 一键预设模型 — 15+ 预设模型（DeepSeek V4、GPT-4.1、Gemini 2.5、Qwen、GLM、Kimi、Groq、Ollama…）
• 🔐 安全密钥管理 — API Key 存入 VS Code Secret Storage，永不记录日志，配置页自动脱敏
• 📊 用量统计 — 状态栏显示端口、余额、上下文用量、请求数、错误数；配置页显示模型级延迟和 token 用量
• 🤖 Explore 子代理 — 自动将 Copilot Explore 子代理路由到你的 OAI 模型，避免回退到 GPT-4.1
• 🖼️ Vision 代理 — 文本模型也可以处理图片，自动用另一个视觉模型描述图片内容