Frequently Asked Questions

Admesh.ai is a large-model aggregation gateway. It wraps mainstream models such as Claude, GPT, DeepSeek, Qwen, GLM, Kimi, Gemini, and MiniMax into a unified OpenAI-compatible interface. All you need is a base URL and an API key to switch between them.

Typically, you only need to make three changes: the Base URL, the API Key, and the model name. The existing OpenAI SDK’s calling methods, message structures, and streaming logic do not need to be modified. If you’re using Anthropic’s native protocol, we also provide compatible endpoints.

https://api.admesh.ai/v1. All models share this single endpoint; the specific model is selected via the `model` field in the request body. Model names follow the format `namespace/model`, such as `anthropic/…` or `openai/…`.

Any OpenAI-compatible client that allows you to customize the Base URL will work, including LangChain, LlamaIndex, Cursor, Cherry Studio, Claude Code, Codex CLI, and others. The documentation site provides step-by-step integration instructions for each client.

Refer to the console's model directory or the public /v1/models API for the current list of available models and their prices. We recommend that you do not hard-code model names in any code; instead, store them in a configuration file so that you only need to update the configuration when switching models.

No. Disaster recovery switching only occurs within the channel pool of the same `public_model`—if you request Claude, the response will always be from Claude. Switching between models would directly affect the quality of the output, and that is not a decision the gateway can make for you.

The gateway automatically switches between multiple channels of the same model: connection failures, timeouts, 429 errors, and 5xx errors will all trigger a switch; when consecutive failures reach a threshold, the channel enters a circuit-breaking state, suspends traffic reception, and periodically checks for recovery.

Pricing is strictly based on usage; you are billed for the actual number of tokens used, with no monthly fees or mandatory subscriptions. Prepaid balances remain valid indefinitely and are not reset monthly. See the Pricing page for details.

Failed requests that do not generate upstream output are not billed, including those resulting from parameter errors, authentication failures, insufficient balance, and rate limiting. If a streaming response is interrupted after some content has already been output, the output that has already been generated will be billed accordingly.

By default, prompts and model output bodies are not persisted. Usage logs record only request metadata, the number of tokens, the model, the amount, error codes, and latency, and are used for billing and troubleshooting.

The server only stores the hash value of the Key and the prefix/last four digits for display purposes. The plaintext is displayed only once at the moment of successful creation and cannot be retrieved again afterward. Once a Key is deleted, it becomes invalid within tens of seconds.

No. Upstream credentials are stored only in encrypted storage; they are not logged, do not appear in browser code, and do not appear in any external APIs.