agents-chat
splash image

Multi-user AI Chat with RedwoodSDK RSC and Cloudflare Agents

This is an experimental project, looking at how to live-stream AI responses back to multiple connected clients. All implementations use Cloudflare durable objects and React Server Components (RSC) with RedwoodSDK.

Deployed at https://agents-chat.jldec.me/

  1. RSC Chat - Uses RedwoodSDK realtime websockets

  2. OpenAI Chat - New! Uses OpenAI Agents SDK

  3. Agent Chat - Uses Cloudflare Agents websockets

  4. Agent SDK Chat - Uses Cloudflare Agents AIChatAgent

  5. TinyBase Chat - Uses TinyBase websockets

  6. Agent Agent Chat - Cloudflare Agents with subagents and MCP tools

  7. Time - Bonus - My very first foray into RSCs and RedwoodSDK

Companion repo for blog post: Multi-user AI chat with RedwoodSDK RSC and Cloudflare agents.

Takeaways

TL;DR

  • Cloudflare is a great platform for building multi-agent multi-user streaming UX.
  • Current UI tooling is immature e.g. with limiting assumptions which block many-to-many connections.

RedwoodSDK RSC:

  • React Server Components provide succinct way to render JSX on the server.
  • RedwoodSDK makes using react with Cloudflare workers easy.
  • The realtime feature of RedwoodSDK pushes updates to connected clients over websockets.
  • Had to disable ssr for hooks (e.g. useChat) not designed to run server-side.

Cloudflare Agents

  • Cloudflare durable objects (DOs) are ideal for agents and subagents. A model with function calling is used for reliable tool and subagent support in Agent Agent Chat.
  • The agents library wraps DOs with additional classes and react hooks.
  • It depends on Vercel's AI SDK for model portabilty, message types, streaming, and tool calling.
  • useChat (AI sdk) and AIChatAgent result in tight coupling between Agents and UI.
  • More work is required to make subagents first-class. The libraries assume single-user streaming, and perform message parsing client-side. The agent-agent implementation has to work around this. See agents #366 and #368.
  • RSC support in Vercel's AI SDK is limited.

OpenAI Agents SDK

  • OpenAI Agents SDK is less mature and focuses on model APIs, not UI integration.
  • It offers APIs for realtime, handoffs, and subagents - not used in this project yet.
  • The stateful Agent abstraction assumes long-running server processes with a single conversation per agent.

TinyBase:

  • DB sync engines can improve UX with local-first client-side persistence.
  • The approach requires careful validation since database operations run on the client.
  • The architecture assumes that connected clients sync the same data.

Links