ocincc — OpenCode Zen free models inside Claude Code

Run OpenCode Zen's free AI models (Big Pickle, Ox Alpha, MiMo, Hy3, Nemotron, DeepSeek Flash, Laguna…) inside Claude Code through a local protocol-translating proxy.

npx ocincc@latest        # from the npm registry
# or, directly from GitHub:
npx github:farseenmanekhan1232/ocincc

That's it. The installer sets up a local LiteLLM proxy and a zen-claude launcher. Your existing claude command and ~/.claude/settings.json are never touched.

Why this exists

Claude Code only speaks the Anthropic Messages API (/v1/messages). Most of Zen's free models are only served via the OpenAI chat/completions format, so pointing Claude Code directly at https://opencode.ai/zen fails for them.

ocincc runs a tiny local proxy that:

Claude Code ──(Anthropic /v1/messages)──▶ localhost:4000 ──translate──▶ https://opencode.ai/zen/v1/chat/completions

What you get

• ✅ All current free Zen models in Claude Code (auto-discovered at launch)
• ✅ Interactive model picker every session (or --model <id>)
• ✅ Context window auto-configured per model (CLAUDE_CODE_MAX_CONTEXT_TOKENS) — 1M-context models get their full window
• ✅ Automatic 429 fallbacks between models
• ✅ Fresh client headers per session so you get the standard free-tier allowance
• ✅ Zero changes to your Claude Code config

Requirements

• macOS or Linux
• Claude Code (claude on PATH)
• Python 3 + curl
• A free OpenCode API key from opencode.ai/auth — no credit card needed for free models

Install & use

npx ocincc@latest          # one-time setup (prompts for your API key)
zen-claude                 # pick a model → start working

Launcher commands:

| Command | Description | |---|---| | zen-claude | Interactive model picker, then launches Claude Code | | zen-claude --model big-pickle | Skip the picker | | zen-claude --continue | Any Claude Code flags pass through | | zen-claude --stop-proxy | Stop the background proxy | | zen-claude --logs | Tail proxy logs |

Everything lives under ~/.config/zen-claude/:

zen-claude           launcher script
.env                 ZEN_API_KEY (chmod 600)
litellm-config.yaml  regenerated fresh on every launch
cache/               version + model-list caches
proxy.log / proxy.pid / venv/

Free models

The model list is fetched live from https://opencode.ai/zen/v1/models on every launch and filtered automatically (*free*, plus stealth freebies like big-pickle and x-preview-f-free). When Zen rotates its catalog, you don't have to change anything.

Current lineup at time of writing:

| Model | ID | Notes | |---|---|---| | Big Pickle | big-pickle | Stealth model, best tool-calling track record | | Ox Alpha Free | x-preview-f-free | Stealth, zero-retention policy | | MiMo-V2.5 Free | mimo-v2.5-free | | | Hy3 Free | hy3-free | | | Nemotron 3 Ultra Free | nemotron-3-ultra-free | NVIDIA trial endpoints | | Nemotron 3.5 Lightning Free | nemotron-3.5-lightning-free | | | DeepSeek V4 Flash Free | deepseek-v4-flash-free | | | Laguna S 2.1 Free | laguna-s-2.1-free | |

How it works (the interesting part)

While building this we read Zen's server source (anomalyco/opencode, packages/console/app/src/routes/zen/) and found:

1. Free-model rate limiting is per IP address, not per API key (ipRateLimiter.ts). Your key just authenticates; everyone behind your IP shares a daily bucket.
2. A header check decides which allowance you get. Requests that look like a genuine OpenCode client get the standard daily limit; anything else gets a stricter fallback limit (FreeUsageLimitError). The check is a substring match on headers such as:

   User-Agent: opencode/<version> (<platform> <kernel>; <arch>)
   x-opencode-client: tui
   x-opencode-project / -session: any stable string

   The launcher sends exactly this shape — version resolved live from npm, IDs generated fresh each session.
3. LiteLLM's /v1/messages endpoint normally bridges OpenAI-provider models to the Responses API; setting LITELLM_USE_CHAT_COMPLETIONS_URL_FOR_ANTHROPIC_MESSAGES=True forces proper chat/completions translation instead.

None of this is spoofing authentication — your requests still use your own API key against the public API; the headers simply identify your client the same way any OpenCode user's would.

Limits & caveats

• Daily quota is real: agentic sessions burn requests quickly; expect occasional 429 Rate limit exceeded errors. Resets daily (UTC). The proxy retries and falls back across models where buckets allow.
• Privacy: most free models may use prompts to improve their models (exception: x-preview-f-free has a zero-retention policy). Don't send sensitive code or secrets.
• Tool-calling quality varies by model — some handle Claude Code's complex tool schemas better than others.
• Free models rotate over time; the tool adapts automatically.
• Not affiliated with OpenCode or Anthropic.

Uninstall

zen-claude --stop-proxy
rm -rf ~/.config/zen-claude /usr/local/bin/zen-claude

License

MIT