ChatGPT Apps SDK: Build an App That Runs Inside ChatGPT

Most "AI app" tutorials stop at a chat UI you host yourself. ChatGPT Apps flip that: the product surface is inside ChatGPT. The user never leaves the place they already talk to the model. That sounds convenient until you hit the real work — architecture, local testing, and the tunnel that makes ChatGPT able to reach your machine.

I walked the official OpenAI Apps SDK quickstart end to end. Not because Todo lists are interesting. Because the wiring is the product.

In the video#

00:00 — Why apps inside ChatGPT are a different surface
00:48 — ChatGPT app architecture
05:18 — Building the Todo app
11:03 — Testing locally
14:31 — Exposing with ngrok
16:00 — Connecting it to ChatGPT
19:23 — Wrap-up

Resources:

Quickstart: OpenAI Apps SDK
Developer mode: platform.openai.com docs
MCP tooling: mcpjam.com

The loop that actually matters#

The Todo app is the vehicle. The loop is the lesson:

Run it locally — prove the app logic without ChatGPT in the path.
Expose it — ngrok (or equivalent) so ChatGPT can call your endpoint.
Register and connect — ChatGPT only talks to what it can reach.

Most people stall on step two and blame the SDK. Fix the tunnel first. Then debug the app. Mixing "is my code wrong?" with "can ChatGPT see my laptop?" burns an afternoon for nothing.

Architecture before vibes#

The video starts with architecture on purpose. You're not pasting a system prompt into a playground. You're shipping something that lives in OpenAI's client — which means auth, developer mode, and how the app talks back into the conversation all need to be clear before you write the Todo handlers.

If you skip that and jump to UI, you'll rebuild the same confusion twice: once locally, once after ChatGPT can't reach you.

What I'd do with a free hour#

Build the tiniest app that completes the full loop once. Local run. Public tunnel. Connected in ChatGPT. Confirm a round-trip.

After you've seen the wiring, the second app is design work — features, UX, what belongs in ChatGPT vs. what belongs in your own product. The first app is ceremony you only want to learn once.

Takeaway#

ChatGPT Apps aren't "prompts with a coat of paint." They're a deployment target. Treat the quickstart like infrastructure practice: architecture → local → tunnel → connect. The Todo list is disposable. The loop isn't.