OpenCode Go Responses Proxy

A lightweight Node.js gateway that translates the OpenAI Responses API to the OpenCode Go chat/completions endpoint. Supports streaming and non-streaming requests, tool calling, reasoning content, and multi-model routing.

What it does

• Exposes /v1/responses (OpenAI Responses API format)
• Forwards to OpenCode Go's /chat/completions endpoint
• Handles SSE streaming with automatic reconnect and client disconnect abort
• Supports function calling, reasoning text, refusals, and image inputs
• Enforces per-model capability limits (tools, reasoning, max tokens, param allowlists)
• Maintains conversation threads via previous_response_id for multi-turn chat

Requirements

• Node.js 20+ (for native fetch and AbortController)
• An OpenCode Go API key in ~/.local/share/opencode/auth.json or via OPENCODE_GO_API_KEY

Quick start

# Run the proxy
node opencode-gateway.mjs

# Proxy listens on http://127.0.0.1:4141 by default

Environment variables

Variable	Default	Description
OPENCODE_GO_API_KEY	—	API key (overrides auth file)
OPENCODE_GO_MODEL	kimi-k2.6	Default model when client omits model
OPENCODE_GO_BASE_URL	https://opencode.ai/zen/go/v1	Upstream base URL
KIMI_PROXY_HOST	127.0.0.1	Proxy bind address
KIMI_PROXY_PORT	4141	Proxy port
KIMI_PROXY_UPSTREAM_TIMEOUT_MS	120000	Upstream request timeout (ms)
OPENCODE_GO_MODEL_CAPABILITIES	—	Override model capability map (JSON)

Minimal Codex config snippet

Add these lines to your existing ~/.codex/config.toml:

model = "kimi-k2.6"
model_provider = "opencode-kimi-proxy"

[model_providers.opencode-kimi-proxy]
base_url = "http://127.0.0.1:4141/v1"
wire_api = "responses"

That is the minimum required. The rest of the settings below are optional tuning.

Full Codex CLI configuration

Create ~/.codex/config.toml (or pass --config):

model = "kimi-k2.6"
model_provider = "opencode-kimi-proxy"
model_context_window = 262144

# Leave room for Kimi's 65,536-token output ceiling
model_auto_compact_token_limit = 196608

[model_providers.opencode-kimi-proxy]
name = "OpenCode Go Kimi Proxy"
base_url = "http://127.0.0.1:4141/v1"
env_key = "KIMI_PROXY_KEY"
wire_api = "responses"
request_max_retries = 2
stream_max_retries = 6
stream_idle_timeout_ms = 300000

Run Codex:

codex

Running proxy + Codex in one command

Use a subshell or process manager. With bash:

(
  node opencode-gateway.mjs &
  PROXY_PID=$!
  sleep 2
  codex "$@"
  EXIT=$?
  kill $PROXY_PID 2>/dev/null
  exit $EXIT
)

Or with concurrently / npm-run-all (if you prefer):

{
  "scripts": {
    "proxy": "node opencode-gateway.mjs",
    "codex": "codex",
    "dev": "concurrently \"npm run proxy\" \"sleep 2 && npm run codex -- $@\""
  }
}

Model capabilities

The proxy ships with a built-in capability map. You can override it entirely via OPENCODE_GO_MODEL_CAPABILITIES:

OPENCODE_GO_MODEL_CAPABILITIES='{
  "kimi-k2.6": {
    "tools": true,
    "reasoning": true,
    "vision": true,
    "json_mode": true,
    "max_output_tokens": 65536,
    "context_window": 262144,
    "supported_params": ["temperature","top_p","presence_penalty","frequency_penalty","max_tokens","stop","seed","n","response_format"]
  }
}' node opencode-gateway.mjs

Built-in models:

Model	Context Window	Max Output	Input $/1M	Output $/1M	Tools	Reasoning	Vision
kimi-k2.6	262,144	65,536	$0.32	$1.34	✓	✓	✓
kimi-k2.5	262,144	65,536	$0.60	$3.00	✓	✓	✓
glm-5	204,800	131,072	$1.00	$3.20	✓	✓	—
glm-5.1	204,800	131,072	$1.40	$4.40	✓	✓	—
qwen3.5-plus	262,144	65,536	$0.20	$1.20	✓	✓	✓
qwen3.6-plus	262,144	65,536	$0.50	$3.00	✓	✓	✓
mimo-v2-pro	1,048,576	64,000	$1.00	$3.00	✓	✓	—
mimo-v2-omni	262,144	64,000	$0.40	$2.00	✓	✓	✓
minimax-m2.5	204,800	65,536	$0.30	$1.20	✓	✓	—
minimax-m2.7	204,800	131,072	$0.30	$1.20	✓	✓	—

Notes:

• Vision support includes image and video input on Kimi and Qwen models; MiMo V2 Omni also supports audio and PDF input.
• MiMo V2 Pro has experimental over-200K pricing: input $2.00, output $6.00 per 1M tokens.
• MiMo V2 Omni and MiMo V2 Pro offer low, medium, and high reasoning-effort variants.
• Token limits are enforced by the proxy's capability map. You can override any model's limits via OPENCODE_GO_MODEL_CAPABILITIES.
• Unknown models fall back to conservative defaults (no tools, no reasoning, no vision).

Endpoints

Method	Path	Description
GET	/healthz	Health check + upstream config
GET	/v1/models	List available models from capability map
POST	/v1/responses	OpenAI Responses API

Architecture

┌─────────┐     /v1/responses      ┌──────────────┐     /chat/completions     ┌─────────────┐
│  Codex  │ ──────────────────────> │  This proxy  │ ────────────────────────> │ OpenCode Go │
│  (CLI)  │  (Responses API format) │  (Gateway)   │  (chat/completions fmt) │  (Upstream) │
└─────────┘                         └──────────────┘                         └─────────────┘

The proxy:

1. Receives OpenAI Responses API requests (input, tools, previous_response_id, etc.)
2. Normalizes content and rebuilds messages as chat/completions format
3. Enforces model capability limits (strips unsupported params, caps tokens)
4. Forwards to upstream and translates the response back to Responses API events
5. Stores conversation threads for previous_response_id continuity

Notes

• The upstream /v1/models endpoint on OpenCode Go returns 404, so the proxy uses a static capability map instead of discovery.
• Reasoning content is cached per tool-call ID for multi-turn tool loops.
• Conversation threads are stored in-memory (LRU, max 100). Restarting the proxy loses thread history.