GHCP2OpenCode

Ollama-compatible API proxy that routes VS Copilot Chat requests to the OpenCode Go subscription API.

When you connect VS Copilot to this proxy, you get access to 14 Go models — including DeepSeek, Qwen, Kimi, MiniMax, GLM, and MiMo — with full agent mode (tool calling) support.

Tested with Visual Studio 2026

Quick Start

1. Install Bun

2. Set your API key

   # .env
   OPENCODE_API_KEY=your-key

3. Run the proxy

   bun run start
   # or on Windows: start.cmd

4. Configure VS Code — set github.copilot.chat.advanced.debug.overrideProxyUrl to http://localhost:11434 and overrideEngineUrl to http://localhost:11434/v1

Connect via Ollama

You can register the proxy as an Ollama instance using the project port:

// settings.json
"ollama.endpoint": "http://localhost:11434"

Endpoints

Endpoint	Format	Used By
/api/tags	Ollama	Copilot model discovery
/api/chat	Ollama	Legacy chat
/api/generate	Ollama	Legacy completion
/api/show	Ollama	Model info
/v1/chat/completions	OpenAI	Copilot chat (primary)
/v1/models	OpenAI	Model listing

Models

All models support tool calling (agent mode). Vision-capable models are listed below.

Model	Vision
DeepSeek V4 Flash
Qwen3.5 Plus	✓
Qwen3.6 Plus	✓
MiniMax M2.5
MiniMax M2.7
Kimi K2.5	✓
Kimi K2.6	✓
GLM-5
GLM-5.1
MiMo V2 Omni	✓
MiMo V2.5	✓
MiMo V2 Pro	✓
MiMo V2.5 Pro	✓

Configuration

Variable	Default
OPENCODE_API_KEY	(required)
SERVER_PORT	11434
SERVER_HOST	127.0.0.1
DEFAULT_MODEL	deepseek-v4-flash

How It Works

1. VS Copilot sends model requests in Ollama format (/api/tags)
2. Proxy forwards chat to the OpenCode Go API (/zen/go/v1/chat/completions)
3. Responses are streamed back in SSE format with proper tool call normalization
4. DeepSeek reasoning_content is cached and re-injected since VS strips non-standard OpenAI fields

Tech Stack

• Bun runtime
• Hono web framework
• Direct fetch to OpenCode Go API (no AI SDK)