I am a solo founder, and Cloudflare is the piece of infra I never have to think about — which is the highest compliment I can give infra. It sits in front of everything at FetchSandbox: TLS, caching the docs for 60+ APIs so they load fast worldwide, and shielding a sandbox API that spins up real containers on demand from abuse. Setup was minutes, not a sprint, and it's been invisible-in-the-good-way ever since. For a one-person team, that reliability is worth more than any single feature.

FetchSandbox runs entirely on AWS. Every sandbox, every spec's engine, and every reproduce→prove run — where we spin up two containers to prove a bug is real, then tear them down seconds later — happens here. As a solo founder I started on Lightsail for dead-simple, predictable hosting, and the best part is knowing the rest of AWS is right there when I need to scale, without a migration. It's the boring-in-the-best-way foundation: reliable enough that I think about my product, not my servers.

I run GPT alongside Claude in FetchSandbox's generation pipeline — parsing OpenAPI specs, discovering realistic workflows, sanity-checking output across 60+ APIs. Where GPT earns its place is breadth and reliability: it's seen every API convention under the sun, and its structured output (function calling / JSON mode) parses cleanly every time — which matters when you're generating at volume and can't babysit each result. It's the dependable workhorse: fast, consistent, and it just works. For a solo founder shipping a lot, that consistency is gold.

I'm a solo founder, and Claude is on both sides of my company. I build with Claude Code — it's written a real share of FetchSandbox, from the sandbox engine to the deploy scripts — and Claude runs inside the product: it distills real API failures into runnable bug patterns, writes the docs for 60+ API specs, and powers the loop that reproduces a bug and reasons about the fix.

What stands out after months of daily use is that the reasoning holds up on hard, multi-step problems where other models hand-wave. Opus for the deep work, Haiku for the fast paths. And because it speaks MCP, my users' agents call my tools through Claude directly — no glue code. As a one-person team, it's the closest thing I have to a second engineer.