zen-proxy

Run Claude Code on OpenCode Zen's free models. A zero-dependency local proxy that speaks Anthropic's Messages API on the front and OpenAI's Chat Completions on the back — so claude never knows it isn't talking to Anthropic.

Claude Code ──POST /v1/messages──► zen-proxy ──POST /chat/completions──► OpenCode Zen
                                   Anthropic ⇄ OpenAI
                                   streaming · tool use · thinking

No API key required. No opencode install required — just Node.

---

Why

Claude Code only speaks the Anthropic API. OpenCode Zen only speaks OpenAI Chat Completions. Left alone you pick one: pay for Anthropic tokens, or give up Claude Code's agent loop.

zen-proxy removes the choice. It terminates Anthropic-format requests locally, rewrites them to OpenAI format, forwards to Zen, then rewrites the response back — translating streaming deltas, tool calls, and reasoning blocks in both directions.

Quick start

git clone https://github.com/Parithosh-Varma/zen-proxy.git
cd zen-proxy
mkdir -p ~/.zen-proxy && cp proxy.js zen-claude ~/.zen-proxy/ && chmod +x ~/.zen-proxy/zen-claude

# start the proxy (defaults: port 8083, model deepseek-v4-flash-free)
node ~/.zen-proxy/proxy.js &

Then run Claude Code through the launcher — it exports the env vars for you, so ~/.claude/settings.json stays untouched:

ZEN_MODEL=hy3-free ~/.zen-proxy/zen-claude -p "Reply with exactly: ZEN-OK"

Prefer a config file? Add this to ~/.claude/settings.json:

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://127.0.0.1:8083",
    "ANTHROPIC_AUTH_TOKEN": "zen",
    "ANTHROPIC_MODEL": "hy3-free",
    "ANTHROPIC_SMALL_FAST_MODEL": "hy3-free"
  }
}

Prerequisites

• Claude Code v2.1.233+ — claude on your $PATH
• Node.js 18+ (tested on Node 20)
• Nothing else. No npm install, no lockfile, no build step.

Why zero dependencies

proxy.js is a single 400-line file using only Node's built-in http and the global fetch. There is no dependency tree to audit, pin, or break on you — which matters when the thing you're running is a credential-adjacent proxy sitting between your editor and a model endpoint.

Configuration

Variable	Default	Description
PORT	8083	Listen port, bound to 127.0.0.1
MODEL	deepseek-v4-flash-free	Zen model id to forward to
ZEN_URL	https://opencode.ai/zen/v1	Upstream endpoint
ZEN_KEY	public	Bearer token — public is the free tier

The zen-claude launcher reads ZEN_PORT, ZEN_MODEL, and ZEN_CONTEXT instead, and defaults ZEN_MODEL to hy3-free.

PORT=9099 MODEL=nemotron-3-ultra-free node ~/.zen-proxy/proxy.js

Endpoints

Method	Path	Purpose
POST	/v1/messages	Anthropic Messages → translated, forwarded
POST	/v1/messages/count_tokens	Local estimate (chars / 4)
GET	/v1/models	Returns the configured model
GET	/health	Liveness check

What gets translated

Anthropic	OpenAI	Notes
system prompt	system message	Flattened from content blocks
tool_use block	tool_calls entry	Arguments streamed as input_json_delta
tool_result block	tool message	Matched back by tool_use_id
thinking block	reasoning_content	Streamed as thinking_delta
stop_reason	finish_reason	tool_use ⇄ tool_calls, max_tokens ⇄ length

Streaming is translated incrementally — the proxy rewrites Server-Sent Events as they pass through, so Claude Code's UI updates live rather than after the full completion.

Rate limiting

Zen's free tier is quota-limited. The proxy stamps every upstream request with a fresh session id (zen-proxy-<pid>-<timestamp>), so its quota bucket is separate from any other OpenCode sessions you're running — requests here won't starve them and vice versa.

If you hit 429 FreeUsageLimitError, swap MODEL to another free-tier id:

hy3-free · deepseek-v4-flash-free · nemotron-3-ultra-free

Security notes

• Binds 127.0.0.1 only — not reachable from the network.
• The bearer token sent upstream is whatever ZEN_KEY holds; the default public is the unauthenticated free tier.
• Claude Code still runs entirely locally. This proxy changes where completions are billed, not where your code is executed.

License

MIT