opencode-go-claude-proxy

Use your OpenCode Go subscription with Claude Code CLI.

Claude Code CLI  ─── Anthropic format ──▶  opencode-go-proxy :3456  ─── OpenAI format ──▶  OpenCode Go
                 ◀── Anthropic format ───                            ◀── OpenAI format ────

Claude Code thinks it's talking to Anthropic. Your requests silently route through OpenCode Go's affordable open models instead.

---

Quickstart (choose one method)

Method 1 — Docker (local build)

1. Clone the repository:

   git clone https://github.com/ahsanjamee/opencode-go-claude-proxy.git
   cd opencode-go-claude-proxy

2. Build the image:

   docker build -t opencode-proxy .

3. Run the proxy (passing your API key):

   docker run -d --name proxy -p 3456:3456 -e OPENCODE_API_KEY=sk-YOUR-KEY opencode-proxy

---

Method 2 — Run from Source (Node.js)

1. Clone the repository and install dependencies:

   git clone https://github.com/ahsanjamee/opencode-go-claude-proxy.git
   cd opencode-go-claude-proxy
   npm install

2. Create an .env file from the example:

   cp .env.example .env

   Edit .env and add your OPENCODE_API_KEY.

3. Start the proxy:

   # For development (hot-reload):
   npm run dev
   
   # For production:
   npm run build
   npm start

---

Step 2 — Configure Claude Code to use the proxy

Create or edit the Claude Code settings file (usually at ~/.claude/settings.json) to point it to the proxy:

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:3456",
    "ANTHROPIC_AUTH_TOKEN": "sk-test",
    "ANTHROPIC_API_KEY": "",
    "ANTHROPIC_MODEL": "deepseek-v4-pro"
  },
  "theme": "dark-ansi",
  "effortLevel": "high"
}

Tip: You can change ANTHROPIC_MODEL to any model listed in the Supported Models table (like minimax-m2.7 or deepseek-v4-pro). The proxy will automatically route your requests to the correct backend.

---

All CLI flags

Flag	Environment variable	Default	Description
--api-key <key>	OPENCODE_API_KEY	—	Required. Your OpenCode Go API key
--port <number>	PORT	3456	Port to listen on
--base-url <url>	PROXY_BASE_URL	https://opencode.ai/zen/go/v1	Upstream base URL
--timeout <ms>	PROXY_TIMEOUT_MS	60000	Upstream idle timeout (resets on each chunk for streams)
--config <path>	PROXY_CONFIG_PATH	—	Path to config.json
--help	—	—	Show help

---

Supported models

Model	Backend
kimi-k2.6, kimi-k2.5	OpenAI compat
glm-5.1, glm-5	OpenAI compat (thinking)
deepseek-v4-pro, deepseek-v4-flash	OpenAI compat (thinking)
qwen3.6-plus, qwen3.5-plus	Alibaba compat
mimo-v2-pro, mimo-v2-omni	OpenAI compat
minimax-m2.7, minimax-m2.5	Anthropic native (long context)

---

API endpoints

Method	Path	Description
POST	/v1/messages	Main chat endpoint
POST	/v1/messages/count_tokens	Token count estimate
GET	/v1/models	Available model list
GET	/health	Health check

---

Troubleshooting

Error: OPENCODE_API_KEY is required
→ The key must be passed as --api-key sk-... or via the OPENCODE_API_KEY env var.
→ When using npm run start, you must use -- --api-key sk-... (double dash before flags).

Tools/skills not working in Claude Code
→ Ensure you're on v1.1.0+. Earlier versions had a wrong SSE delta type that broke all tools.

Reasoning text showing in Claude Code output
→ Ensure you're on v1.1.0+. Reasoning is now properly emitted as hidden thinking blocks.

400 errors from beta headers
→ Set CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 before running claude.

Stream times out / AbortError during long agent runs
→ The proxy uses an idle timeout for streaming: the timer resets every time a chunk arrives from the model. If no data arrives for PROXY_TIMEOUT_MS (default 60 s), the connection is aborted.
→ For long agent tasks (e.g. Kimi k2.6 reasoning for minutes), increase the timeout in your .env:

PROXY_TIMEOUT_MS=300000   # 5 minutes

→ The proxy also sends SSE keepalive comments every 3 s to keep the Claude Code side alive, but this does not affect the upstream timeout.

---

What gets translated

Anthropic → OpenAI (outgoing request)

Anthropic	OpenAI
system string/block array	messages[0] with role: "system"
Text content blocks	content string
thinking blocks	reasoning_content
tool_use blocks	tool_calls
tool_result blocks	messages with role: "tool"

OpenAI → Anthropic (incoming response)

OpenAI	Anthropic
finish_reason: "stop"	stop_reason: "end_turn"
finish_reason: "tool_calls"	stop_reason: "tool_use"
finish_reason: "length"	stop_reason: "max_tokens"
delta.content text	text_delta in content_block_delta
delta.reasoning_content	thinking_delta in content_block_delta
delta.tool_calls[].function.arguments	input_json_delta in content_block_delta

---

Automatic Model Mapping

Claude Code internally uses claude-* models (like claude-haiku-* for background tasks). OpenCode Go doesn't natively support these IDs.

The proxy automatically maps all Claude models to their closest OpenCode equivalents:

• claude-haiku-* → qwen3.5-plus
• claude-sonnet-* → qwen3.6-plus
• claude-opus-* → qwen3.6-plus

If Claude Code requests an entirely unknown model, it safely falls back to qwen3.6-plus to prevent 401/400 crashes.

Customizing aliases

You can override these mappings by editing ~/.opencode-proxy/config.json:

{
  "modelAliases": {
    "claude-haiku-*": "kimi-k2.5",
    "claude-sonnet-*": "deepseek-v4-pro"
  },
  "global": {
    "defaultModel": "kimi-k2.5"
  }
}

---

Disclaimer

This project is an unofficial community tool and is not affiliated with, endorsed by, or associated with Anthropic or OpenCode Go. Use it at your own risk.

Credits

Created by ahsanjamee.

---

License

MIT