Relay
Relay is the reliable delivery layer for LLM APIs
If you want to be notified when Relay launches on Product Hunt June 17, follow us here: https://www.producthunt.com/products/relay-17
Hey IH,
I've shipped four AI products solo over the last two years (most recently MathLab PvP and EthnoTrace AI). Every single one needed reliability code for the LLM API - retry, backoff, failover when Anthropic 529s, caching identical prompts. I wrote that loop four times, each time with a different bug. The fourth time I admitted it: this is a product, not a yak to shave.
So I'm building Relay -import { Relay } from "@relay-api/sdk" and you get auto retry with decorrelated jitter, automatic provider failover (Anthropic ↔ OpenAI), smart caching, and a live dashboard. Drop-in replacement for @anthropic-ai/sdk - same API, just keeps working when the provider doesn't.
Think Resend, but for LLM APIs. Same shape of problem - every builder rewrites the client code in every project, badly. Resend collapsed it for email; Relay does it for AI.
Where I am right now (week 1, validation phase):
🟢 Landing live: relay-sdk.vercel.app (still on a Vercel subdomain — holding the domain $ until validation lands)
🟢 Waitlist with Supabase + Resend welcome email
🟢 SDK skeleton open-sourced: github.com/relay-llm/sdk (TypeScript, MIT, dual ESM/CJS)
🟢 Worker scaffolded - Hono + Cloudflare Workers + KV cache + retry/failover router
🟡 Worker not deployed yet - week 2 task
🟡 Dashboard placeholder only - week 4
🟡 No customers, no payments - Paddle will be week 4
What I'm looking for from you:
Devs who've felt LLM API reliability pain - reply or DM me. 15 min discovery call, I want to hear the specific bug that ate your evening. I'll trade free 6 months of Pro at launch.
Feedback on the positioning - does "Resend for AI agents" land, or does the analogy feel stretched?
Weekly updates from here. Next milestone: Worker actually deployed, first reliable proxy call goes through (week 2).
Roast / ask anything. 🇰🇿
— Alikhan
Comment