OCFreeRelay

English | 简体中文

Standalone OpenCode free-worker LLM gateway: an OpenAI-compatible relay that serves OpenCode's free models through your own API keys.

• Accepts OpenAI-compatible client requests (/v1/chat/completions, /v1/models)
• Transparent passthrough to https://opencode.ai/zen/v1 (configurable)
• Free-only models: auto-scrapes the Zen pricing page and serves ONLY free models (list + chat); paid models are never exposed
• Multi-key / multi-account sticky affinity until 429, then rotate + cooldown (keeps prompt cache warm)
• Optional OpenCode CLI identity header synthesis (Cloudflare / VPS)
• Minimal free-model body fixes (strip client_metadata, thinking-model reasoning_content, effort aliases)
• Management page at / for keys, base URL, proxies, and status

Quick start

npm install
npm run build
npm start
# or: npm run dev

Default port: 9876 (override with PORT or admin settings).

• Admin UI: http://127.0.0.1:9876/
• Chat: POST http://127.0.0.1:9876/v1/chat/completions
• Models: GET http://127.0.0.1:9876/v1/models

Configuration

Source	Purpose
Admin UI	Base URL, workers (API keys), proxy pool bindings, CLI header synthesis
data/settings.json	Persisted settings (auto-created)
PORT	Listen port
OCFREERELAY_SETTINGS_PATH	Custom settings file path
OPENCODE_SYNTHESIZE_CLI_HEADERS	true to synthesize CLI identity headers (also toggleable in admin)
OPENCODE_USER_AGENT / OPENCODE_CLIENT / OPENCODE_PROJECT	CLI default values
OCFREERELAY_PRICING_URL	Override the Zen pricing page URL used to scrape free models

Free-only model serving

This gateway serves only free models — paid models are never exposed to clients.

• On boot it scrapes the OpenCode Zen pricing page (https://opencode.ai/docs/zen) and keeps every model whose Input & Output price is Free.
• GET /v1/models returns only those free models (upstream's paid entries are dropped).
• POST /v1/chat/completions rejects any request for a non-free model with 403 model_not_allowed before any upstream call.
• The scraped set is cached to data/free-models.json. If a scrape fails, the last successful set is kept; before the first successful scrape a static baseline of the currently-known free ids is used.

Current free models:

big-pickle  deepseek-v4-flash-free  mimo-v2.5-free  laguna-s-2.1-free
ling-3.0-flash-free  longcat-2.0-free  north-mini-code-free  nemotron-3-ultra-free

Proxy pool (IP isolation for OpenCode free)

OpenCode free accounts are often IP-limited. Bind each worker to a different pool proxy:

1. Manual — Admin → Proxy pool → add HTTP/SOCKS5 host:port
2. Clash subscription — add subscription URL → 拉取 (fetch)
   • Tries multiple User-Agents (clash first). Some providers return full YAML only for the clash UA; other UAs return base64 vless:// lists.
   • Imports http/socks (direct) and vless/hysteria2/tuic/… (via Clash bridge)
3. Clash bridge — for protocol nodes:
   • Run Mihomo/Clash Meta locally with the same subscription
   • Enable bridge in Admin: controller http://127.0.0.1:9090, mixed-port (often 7892), selector group 主代理
   • Gateway switches the select-group per worker, then exits via local HTTP proxy
4. Bind — each Worker selects a pool node via proxyId

Tests

npm test

License

MIT

Community

• Linux.do — open-source & developer community