Build ChatGPT App

OverviewInstructionBenchmarkShowcase

Original OpenAI Skill · independently packaged

Turn a request into an app that lives inside the chat.

Build ChatGPT App guides a coding assistant through the Apps SDK in a set order: read the current docs first, choose one app shape, plan the tools before writing code, then scaffold an MCP server and widget UI. At the end it reports exactly which validation actually ran.

Capability overview0:08 · illustrative animation
An 8-second capability overview. This is an illustrative animation, not test evidence. Real run evidence is on the Showcase page and the validation boundary is on the Benchmark page.

How a request becomes an app

One message in. Four connected artifacts out.

A ChatGPT app is more than a server or a widget on its own. The value is in the connections: a tool the model can choose correctly, a UI resource the host can render, and a bridge that carries results between them.

Your request

Create a tiny Node and vanilla-widget app named Release Compass, with a show_release tool. Keep the first iteration local.

  1. 1 · App shapevanilla-widgetA small starter where a single HTML widget is enough. The Skill picks one primary archetype and says which one, so you can correct it early.
  2. 2 · Tool contractshow_releaseA “Use this when…” description, an explicit input schema, and honest annotations: read-only, non-destructive, closed-world, idempotent.
  3. 3 · Widget resourceui://widget/main-v1.htmlRegistered with the MCP Apps UI MIME type, a declared CSP and a border preference. It is linked from the tool through _meta.ui.resourceUri.
  4. 4 · Widget, hydrated by the tool result
    Release Compass starterRelease Compass

    This starter uses the MCP Apps bridge first, keeps follow-up messaging on ui/message, and limits window.openai to optional host signals.

    Call show_release from the widgetAsk the host to explain this app

    Schematic of the generated widget’s hydrated state, using its real copy. It is not a screenshot from a ChatGPT host.

  5. 5 · Validation reportWhat ran, and what did not
    Ran: script syntax, four files generated, tool / resource / read-only metadata presentDid not complete: dependency install, compile, local /mcp runtime, live host

This diagram is based on the real Release Compass starter produced by the Skill’s bundled script. The Showcase page holds the actual run record.

The exchange

What you bring, and what comes back.

You bring

  • The use case and the main user flows
  • Which tools only read and which change things
  • Whether it is a demo or a production target, private or public
  • Preferred backend language and UI stack (TS or Python; React or vanilla)
  • Auth needs, external API domains, and hosting plans
  • Your repository, if you are refactoring an existing app

You get back

  • The chosen app archetype and why it fits
  • A tool plan: names, schemas, annotations, outputs
  • A recommended starting point: official example, ext-apps example, or the local fallback scaffold
  • MCP server and widget code, with the MCP Apps bridge first and window.openai as an optional extension
  • A validation report against a minimum working-repo contract
  • Local run, tunnel and ChatGPT Developer Mode setup steps
  • When asked: hosting guidance, a submission checklist, a short stakeholder summary

Five supported shapes

It picks the smallest shape that does the job.

The Skill infers a single primary archetype before choosing examples or scaffolds. It moves to the heavier shapes only when the request calls for them.

  1. 01tool-onlyNo in-chat UI needed: search, fetch, retrieval or background actions. Connector-like apps default to standard search and fetch tools.
  2. 02vanilla-widgetA small demo, workshop starter or simple inline widget. Root-level server plus public/ assets.
  3. 03react-widgetA polished, component-based UI. Split server/ + web/ layout, starting from official OpenAI examples.
  4. 04interactive-decoupledBoards, maps, editors, games and dashboards. Data tools are separated from render tools so the widget stays mounted.
  5. 05submission-readyOnly when you ask for public launch: domain, accurate CSP, review-safe auth and submission artifacts.
First choiceOfficial OpenAI examplesWhen a close example already matches the stack or interaction pattern.
SecondVersion-matched ext-apps examplesFor lower-level or more portable MCP Apps baselines.
FallbackBundled Node scaffoldA tiny Node + vanilla widget starter for when no close example fits.

Environment and dependencies

What it needs to run.

Assistant
A coding assistant that loads Skill folders. Invoke it as build-chatgpt-app. The package includes SKILL.md, seven reference guides, the fallback script and an agents/openai.yaml interface file.
Current docs
The preferred path is the $openai-docs Skill. Otherwise the assistant uses the OpenAI developer docs MCP server (developers.openai.com/mcp, streamable HTTP). If search fails, it fetches the canonical Apps SDK pages by URL.
Repository access
The assistant needs permission to create or edit project files.
Node fallback
Node.js and npm. The generated project declares @modelcontextprotocol/ext-apps, @modelcontextprotocol/sdk, zod, tsx, typescript and @types/node.
Testing in ChatGPT
A public HTTPS URL for the local server (for example via a tunnel) and an account with access to ChatGPT’s developer features. Labels and eligibility are set by the host and can change.

No separate OpenAI API key is needed to generate a local starter. That does not make the whole workflow free. Your assistant’s model usage, ChatGPT account features, tunnelling, hosting and any paid APIs your app calls can each carry their own costs or requirements.

Fit

Who it is for, and where it stops.

A good fit

  • Developers building an MCP-backed ChatGPT app from scratch
  • Teams repairing widget integration or bridge wiring
  • Refactoring a demo toward annotations, CSP and URI versioning
  • Planning tools before code for a stateful or connector-like app

Not what it does

  • It does not grant ChatGPT developer or publishing access
  • It does not host your server or manage your secrets
  • It does not guarantee directory approval
  • It does not add every production control automatically. Auth, permissions and review remain your decisions

What has been verified

Proven so far: package integrity and starter generation.

Static reviewReached for the generated starter
SyntaxScript checked; app not compiled
Local runtimeNot reached: install timed out
Host loopNot run in ChatGPT

All 11 original Skill files match their pinned upstream SHA-256 hashes. The bundled script passed a Node syntax check and generated the four expected starter files. npm dependency installation timed out, so compilation, the local MCP endpoint and live ChatGPT behavior remain unverified. The Benchmark page has the full evidence trail.