Using Ox Alpha in Claude Code: setup guide, limits, and the data risk

On the OpenRouter page tracking Ox Alpha's traffic, Claude Code ranks second, with 1.4 trillion tokens processed. Hundreds of developers have already connected Anthropic's terminal to a model whose operator is officially unknown. The incentive is obvious: a one-million-token context window at a price of zero during the preview. The connection is technically simple. The real question is not how, but at what hidden price.

Ox Alpha is a stealth model that appeared on OpenRouter on August 20, 2026, without any lab claiming authorship. It is built for coding and long-running agentic work — exactly Claude Code's terrain. This guide shows how to point Claude Code at it with four environment variables, what of the original behavior survives, and why you pay the zero price in a different currency: your prompts.

What is Ox Alpha and why does it end up in Claude Code?

Ox Alpha is a reasoning model for long-horizon software engineering, accepting text, image, and video as input. The context window holds 1,048,576 tokens, and the output limit is 131,072 tokens per response. It supports function calling via the tools and tool_choice parameters and accepts response_format for JSON output. During the preview, OpenRouter shows a price of zero, as the specification on the model page confirms.

This combination of huge context, tool calling, and zero cost pushes developers toward command-line agents like Claude Code. The model endures long sessions full of files, logs, and test output without saturating. A snapshot of platform data from August 24 shows a median latency of 5.74 seconds, around 23 tokens per second, and 99.98 percent availability over three days. Solid numbers, but a snapshot, not a guarantee.

Who operates the model remains officially unknown. The forensic fingerprint, from the tokenizer to an error code typical of Z.AI, points to GLM from Zhipu, the Chinese lab behind z.ai. For a while Gemini 3.5 Pro was considered a candidate — a mistake the vision test disproved. This origin matters for the final part of this guide, because it shifts the question of where your data ends up.

How do you connect Ox Alpha to Claude Code?

Claude Code speaks the Anthropic Messages protocol and expects an endpoint that answers under /v1/messages. OpenRouter offers exactly that: an endpoint called Anthropic Skin that accepts requests in the native Anthropic format and forwards them to the provider serving Ox Alpha. No separate proxy is needed.

1. Create a key on OpenRouter

Sign up with OpenRouter and generate an API key in the Keys section. The key has the format sk-or-v1-.... During the preview you need no credit, but you need the key because it identifies your requests.

2. Set the environment variables

Export the variables in the terminal before starting Claude Code. The tricky part is the empty Anthropic key. It must be set as an empty string, otherwise Claude Code tries to authenticate directly with Anthropic and ignores OpenRouter.

export ANTHROPIC_BASE_URL="https://openrouter.ai/api"
export ANTHROPIC_AUTH_TOKEN="sk-or-v1-YOUR_KEY"
export ANTHROPIC_API_KEY=""
export ANTHROPIC_MODEL="stealth/ox-alpha"
export ANTHROPIC_SMALL_FAST_MODEL="stealth/ox-alpha"

The most common source of errors is the base URL. It must be exactly https://openrouter.ai/api, because Claude Code appends /v1/messages itself. If you enter /api/v1, you get a doubled /v1, and every call fails without the error explaining the cause.

3. Start Claude Code

Run the usual command in the same terminal session:

claude

4. Check whether Ox Alpha is active

Type /status or /model in the session. The active model must show stealth/ox-alpha, not a Claude name. If you still see an Anthropic model, ANTHROPIC_API_KEY was almost always not an empty string. Close the session, check the variables, restart.

Does the connection really work?

The legitimate concern is that a foreign model breaks Claude Code's agentic features. The good news: OpenRouter's Anthropic Skin passes through reasoning blocks, native function calling, streaming, and multi-turn context — exactly the mechanisms the agent is built on. The practical proof is the traffic itself, with Claude Code in second place of Ox Alpha usage, right behind Hermes Agent.

• Tool use (tool calling): Works, tools and tool_choice are supported.
• Streaming and multi-turn context: Passed through via the Anthropic Skin.
• Reasoning blocks (thinking): Work when the model generates them.
• Context window: 1,048,576 tokens.
• Official support from Anthropic: No, the integration is not guaranteed.
• Confidentiality of prompts: Prompts and responses are stored.

But there is an honest limit: OpenRouter points out that Claude Code is built around the semantics of Anthropic requests and that the integration is only guaranteed with the first-party provider. Translated, that means the connection works but is not supported. A Claude Code update can break it without warning.

What is the real problem with your data?

This is where it is decided whether you can truly use Ox Alpha without concerns. The technical connection is clean; the data model is not. OpenRouter's policy for stealth models is unambiguous: prompts and responses are stored by the provider and not used for training. The second half is reassuring, the first is not. Everything you send through stays with a provider without a public name, as the data-risk overview from Wavect also emphasizes.

With Claude Code this weighs heavier than elsewhere. Claude Code does not send a single question, it sends your repository. It reads files, attaches them to the context, sends diffs, logs, and sometimes secrets you never wanted to share. With Ox Alpha, all of that ends up in an archive that the fingerprint links to GLM from Zhipu — in all likelihood on Chinese infrastructure outside the European Union.

The timing is remarkable: the free phase presumably ends around August 27, 2026, only about a week after launch, as AI Catchup reports citing OpenCode. So if you want to test, you have a small window.

When should you use Ox Alpha and when not?

You can use it for toy code, personal experiments, already public open-source projects, and throwaway prototypes — everything you would publish anyway. In these cases the free million-token context window is a gift. Enjoy it while the preview lasts.

You should not use it for proprietary code, client repositories, personal data, credentials, and anything covered by a contract or a non-disclosure obligation. In these cases the zero price is a trap, because the real currency is the confidentiality of your work. If you work for multiple clients, keep the environments strictly separated so no client context accidentally lands in an Ox Alpha session.

Is there a more solid alternative?

If the idea of switching off the Anthropic models entirely does not convince you, there is a middle way. The open-source project claude-code-router places itself as a proxy in front of Claude Code and decides, depending on the task type, which endpoint serves the model. On August 24, 2026, it had collected 36,850 stars on GitHub. With it you keep Claude for delicate tasks and only reroute to Ox Alpha what you do not mind sharing, for example boilerplate or test generation.

The router adds an extra component that has to be maintained. But it gives you back control over which part of your work goes where. That is exactly the control you lose when you point the entire terminal at a stealth model across the board.

Key takeaways

• Four environment variables are enough, because OpenRouter already speaks Claude Code's language with the Anthropic Skin. No additional software is needed.
• The Anthropic key must be set as an empty string, not omitted. Otherwise Claude Code keeps talking to Anthropic.
• Tool calling, streaming, and reasoning survive, but the connection is not supported, and an update can break it.
• The zero price is not the real cost. The stealth policy says prompts and responses stay with the provider, and the fingerprint points to GLM from Zhipu outside the EU.
• For client code or confidential data the connection should be avoided, because it becomes a data transfer to a recipient that does not even have a name.

FAQ: frequently asked questions

Conclusion

Try Ox Alpha on Claude Code with a project you could publish tomorrow, not the one you invoice to a client. The connection is clean, and the free one-million-token context window is tempting. But the confidentiality of code does not come back once it has left your machine. This article was editorially reviewed and fact-checked against OpenRouter platform data. You can find more about me and my work on the about page, and further guides in the news overview. For advice on integrating AI models into your development workflow while keeping control of your data, use the contact form.