AgentFlow: Building an AI Agent Distribution Platform from Scratch

I built AgentFlow - a platform that lets you connect your AI agents, workflows, or LLM endpoints, then distribute them across teams and organizations with built-in access control and conversation tracking.

The problem: You’ve built cool AI agents, but how do you actually get them into users’ hands? You need auth, multi-tenancy, usage tracking, conversation history, and a way to swap between different AI backends without rebuilding your frontend.

I’ll walk through the platform architecture:

• Multi-tenant organization system with group-based access control (users → groups → AI models)
• Conversation management with feedback collection
• How any team member can chat with approved AI connections through one interface

Technical deep dive on the Universal Endpoint Router - the core system that lets you plug in ANY AI backend:

• Connect OpenAI, Anthropic, custom webhooks, or your own AI SDK agents
• Automatic stream format conversion (SSE → UI Stream, JSON → Streamed Text)
• Template-based request/response mapping with JSONPath extraction
• Non-blocking persistence that intercepts streams in real-time without affecting latency

Tech stack: Next.js 14, TypeScript, AI SDK 6, Supabase + Row-Level Security, Clerk