Migrating from Official SDKs to a Unified Gateway: Five Steps with Zero Refactoring Cost

1. Extract the Base URL into configuration

If your code still hardcodes https://api.openai.com/v1, the first step isn't to change the address—it's to move it into an environment variable. This determines whether you can roll back at any time.

Base URL and endpoint documentation →

2. Switch to the gateway address and Key

There are only two pieces of information you need to remember for Admesh: the Base URL is https://api.admesh.ai/v1, and the Key is created in the console. The plaintext Key is shown only once at creation; afterward, only the prefix and last four characters are retained.

3. Change model names to namespace/model

The gateway uses public model names in the namespace/model format (for example anthropic/…, openai/…, deepseek/…, qwen/…). The complete list is based on /v1/models or the model catalog in the console; the documentation site also has a [model list](https://docs.admesh.ai/zh/docs/api-reference/models) explanation.

Don't hardcode model names in your code—put them in configuration, so changing models is just a config change.

4. Dual-write comparison

1. Send the same batch of real requests to both the official API and the gateway.
2. Compare three things: the response structure (whether fields match), the usage field (Token counting methodology), and streaming chunk behavior.
3. Check error responses: deliberately send a request with a wrong parameter and confirm the error code format matches expectations.
4. Record the differences in latency and cost between the two sides as a baseline for later optimization.

5. Canary rollout and rollback

• Switch non-critical traffic first (internal tools, low-risk tasks) and observe for a week.
• Monitor three metrics: error rate, P95 latency, and cost per request.
• Keep the rollback plan simple: change the environment variable back to the official address—no code changes needed.
• Once stability is confirmed, migrate your main workload and update the team's documentation and alerting configuration accordingly.

What about the native Anthropic protocol?

Some code directly uses Anthropic's Messages API shape (/v1/messages, required max_tokens, a separate system field). Admesh also provides an [Anthropic-compatible endpoint](https://docs.admesh.ai/zh/docs/api-reference/anthropic-compatible), so these projects don't need to be rewritten in OpenAI style—just change the Base URL and Key.