Support OpenAI ChatKit

[DouweM]

Description

• https://platform.openai.com/docs/guides/chatkit
• https://github.com/openai/chatkit-python/blob/main/docs/server.md

Similar to:

• https://ai.pydantic.dev/ag-ui/
• Support Vercel AI Data Stream Protocol #2923

References

No response

[safina57]

@DouweM I would be happy to start working on this if you don't mind

[safina57]

I have a question about the Integration Strategy

When adding support for OpenAI ChatKit, there are two main approaches:

1. Protocol Compatibility (Agent-First, Pydantic AI Style)

similar to the Vercel one, We implement adapters that translate Pydantic AI agent events into ChatKit-compatible protocol events. Maintaining agent-first architecture (Pydantic AI agents remain the core abstraction)
This approach is lightweight, decoupled, and consistent with Vercel AI/AG-UI adapters.
We do not need to use the official ChatKit Python backend/server or OpenAI SDK directly.

2. Full ChatKit Backend Integration (Server-First, OpenAI SDK Style)

We Wrap or subclass the official ChatKitServer class from the openai-chatkit library.
Implement the respond() method, using Pydantic AI logic inside ChatKit’s server lifecycle.
This approach is tightly coupled to OpenAI’s SDK and backend patterns which enables advanced features: widgets, persistent threads, file attachments, client tools, etc.

If you only need protocol compatibility and want to keep Pydantic AI as the main agent system, we can implement the agent-first adapter approach
If you need full ChatKit backend features (widgets, storage, advanced UI), then we consider wrapping the ChatKit server and integrating Pydantic AI logic inside its respond() method.

The official ChatKit examples and documentation are tightly coupled to OpenAI’s SDK.

[DouweM]

@safina57 Thanks for offering to work on this! Keep an eye out on the refactoring I'm doing in #2923; we should follow the same UIAdapter-based approach for OpenAI ChatKit that we do for AG-UI and Vercel AI. That means option 1 Protocol Compatibility from your proposal. I imagine we can eventually implement the advanced features as well assuming they're part of the public protocol. With AG-UI we already support client tools, so we should be able to make that work here as well.