cmdcode
OpenAI-compatible proxy in front of Command Code's HTTP API. Use your Command Code subscription with any client supporting OpenAI Chat Completions - no vendor lock-in.
What it does
┌──────────────────┐ ┌───────────────┐ ┌─────────────────────────┐
│ OpenCode │ POST │ proxy │ POST │ Command Code API │
│ LiteLLM │ ────> │ :18080 │ ────> │ /alpha/generate │
│ curl / any SDK │ <──── │ (local) │ <──── │ (NDJSON) │
│ │ SSE │ │ NDJSON│ │
└──────────────────┘ └───────────────┘ └─────────────────────────┘
The proxy translates the OpenAI /v1/chat/completions protocol to
Command Code's custom wire format, preserving the exact API fingerprint
(headers, body structure, auth) so the upstream cannot distinguish proxy
traffic from the official cmd CLI.
Quick start
# Install from crates.io (requires Rust 1.75+)
# Or build from source
# Run
# listening on http://127.0.0.1:18080
# Test
Features
- OpenAI-compatible -
/v1/chat/completions(stream + non-stream) - Streaming - full SSE translation (text, reasoning, tool calls)
- Tool calls - OpenAI function calling ↔ Command Code tool-call format
- Rust + Pingora - production-grade, sub-millisecond overhead
- CLI fingerprint - sends the exact headers/body the
cmdCLI sends - Auto-discovery - model catalog parsed from CLI's bundled
models.md - Reasoning effort -
low/medium/high/xhigh/maxsupport - Retry logic - automatic retry on transient upstream failures (502/503/504)
- Health check -
GET /healthfor monitoring - CORS - optional CORS headers for browser clients
- Concurrency - Pingora handles multiple simultaneous agents
Wire it into OpenCode
Add to your opencode.json:
{
"provider": {
"command-code": {
"npm": "@ai-sdk/openai-compatible",
"name": "Command Code",
"options": {
"baseURL": "http://localhost:18080/v1"
},
"models": {
"xiaomi/mimo-v2.5": { "name": "MiMo V2.5", "reasoning": true },
"gpt-5.6-luna": { "name": "GPT-5.6 Luna", "reasoning": true }
}
}
},
"model": "default"
}
Restart OpenCode and you're done.
Wire it into anything
Any OpenAI-compatible client works:
Python:
=
Node.js:
import OpenAI from "openai";
const client = ;
LiteLLM:
model_list:
- model_name: command-code
litellm_params:
model: openai/xiaomi/mimo-v2.5
api_base: http://127.0.0.1:18080
api_key: not-needed
Configuration
| Env var | Default | Description |
|---|---|---|
COMMAND_CODE_PROXY_PORT |
18080 |
Listen port |
COMMAND_CODE_PROXY_HOST |
127.0.0.1 |
Bind address |
COMMAND_CODE_API_BASE |
https://api.commandcode.ai |
Upstream API |
COMMAND_CODE_PROXY_TIMEOUT |
600 |
Upstream timeout (seconds) |
COMMAND_CODE_PROXY_RETRIES |
2 |
Retry count for transient failures |
COMMAND_CODE_PROXY_CORS |
(unset) | CORS origin header |
COMMAND_CODE_PROXY_DEFAULT |
xiaomi/mimo-v2.5 |
Default model |
COMMAND_CODE_PROXY_MODELS |
(unset) | Comma-separated model allowlist |
COMMAND_CODE_PROXY_LOG_FILE |
(unset) | Log file with size-based rotation |
COMMAND_CODE_PROXY_LOG_MAX_BYTES |
52428800 |
Rotate after this many bytes |
COMMAND_CODE_PROXY_LOG_KEEP |
5 |
Rotated log backups to keep |
COMMAND_CODE_PROXY_INCOMING_TOKEN |
(unset) | Require bearer token on API routes |
COMMAND_CODE_PROXY_TLS_CERT |
(unset) | TLS cert path (with KEY enables HTTPS) |
COMMAND_CODE_PROXY_TLS_KEY |
(unset) | TLS key path |
Endpoints
| Method | Path | Description |
|---|---|---|
GET |
/v1/models |
List available models |
GET |
/health |
Health check (status, version, upstream) |
GET |
/metrics |
Prometheus-formatted metrics |
POST |
/v1/chat/completions |
Chat completion (stream + non-stream) |
See docs/setup.md for supervision (scripts/supervise.sh), log rotation,
and soak (scripts/soak.sh) tooling.
Performance
Proxy overhead over the direct upstream path, measured by the through-proxy
benchmark (test_benchmark_through_proxy_vs_direct): 500 requests against the
real Pingora proxy vs. the mock upstream, warm, on loopback.
direct : p50=0.427ms p95=0.469ms p99=0.519ms
proxy : p50=0.804ms p95=0.991ms p99=1.247ms
overhead: p50=+0.377ms p95=+0.522ms p99=+0.728ms
The overhead is sub-millisecond at every percentile. These are machine-specific
warm-loopback numbers for reference; the regression test asserts a relative
bound (p50 < direct + 2ms, p99 < direct + 10ms) so CI noise, which affects
both paths equally, cancels out.
Documentation
- Architecture - wire format translation details
- Setup guide - installation, Docker, nginx
- OpenCode integration - step-by-step wiring
License
MIT