Unified API for
Agentic AI

All AI models and APIs have different interfaces and require different payloads. The Relay helps you standardize the interface and payload as well as handle quirks and edge cases.

While you can also achieve the same by writing your own code, keeping up with the changes in the AI space can be time-consuming and error-prone. Besides, most AI inference APIs are poorly documented and have a steep learning curve.

Since we at ChatBotKit specialize in conversational AI technology, we can help you keep up with the changes and make sure your integration is always running no matter what changes the AI model providers make.

Simply put - we'll handle the hard stuff so you can focus on building great conversational AI experiences.

The most suitable use case for Relay is when you are building prototypes, experiments, MVPs and other small projects.

That being said, Relay can scale to handle large projects as well. We just do not recommend it at this stage. If you are looking for a scaleable solution, we recommend using ChatBotKit directly.

The relay speaks the ChatBotKit Conversation Complete API.

You simply need to send POST requests to https://relay.cbk.ai/api/v1/conversation/complete

For more information on how to use the Conversation Complete API, please refer to the ChatBotKit API Reference.

The relay accepts a subset of that API: model, messages, backstory, functions and extensions. Anything that reaches into a ChatBotKit account - bots, datasets, skillsets, contacts, attachments - is rejected with an explanation rather than quietly ignored, because the relay has no account of yours to reach into.

Within extensions you can use backstory and features. Of the conversation features, the relay supports skills, batch and silent - the three that only shape the prompt. If you need the rest, use ChatBotKit directly.

The Relay API is compatible with ChatBotKit SDK, which provides an extremely lightweight and easy-to-use interface for the API.

You must authenticate your model requests with your own provider API key. The Relay does not provide inference credentials - a request without them is rejected. Each provider has their own way of generating a key and assigning quotas, so please consult with their documentation for more information.

Your key travels in the model configuration, alongside the provider it belongs to:

The provider is required - we will not guess it from the model name, because a wrong guess would send your key to the wrong vendor. Supported providers are bedrock, cloudflare, deepseek, groq, mistral, moonshot, openai, openrouter, perplexity, qwen, vercel, vertex and zai.

You can spell the key authorization= instead of credentials= if you prefer. To point at your own compatible endpoint, add /endpoint=[https_url].

The Relay has no database of its own. It validates your request, forwards it to the ChatBotKit API for processing, and streams the response back. No conversation record is created, and your data is not used for service improvements.

Your provider API key is passed through to the provider you named and is not persisted by the Relay. Requests are not tied to a ChatBotKit user account - which also means that if a call fails we have very little to go on when helping you debug it. We may be able to help you interpret the error message.

Requests are processed under a single shared platform account, and operational error reporting applies as it does to the rest of the platform. If you are handling regulated or otherwise sensitive data, use ChatBotKit directly, where the account, the retention settings and the privacy controls are yours.

Please join our Discord Server for support and community access.