Why You Need a ChatGPT App Framework (April 2026)

ChatGPT Apps are interactive UIs that render directly inside ChatGPT conversations. They’re built on MCP Apps, an official extension to the Model Context Protocol, and they’ve grown fast since launching in January 2026. OpenAI now has a ChatGPT App Directory with 300+ published integrations.

But building a ChatGPT App from scratch still means solving the same infrastructure problems on every project. A ChatGPT App framework changes that.

TL;DR: sunpeak is a ChatGPT App framework built on the MCP Apps protocol. It gives developers a local inspector, UI components, CLI scaffolding, a testing framework with unit/e2e/visual regression/multi-model eval support, and multi-host portability across ChatGPT, Claude, VS Code, and Goose.

What is a ChatGPT App Framework?

A ChatGPT App framework provides the development infrastructure for building apps on top of the MCP Apps protocol. The protocol defines how the application works: the rendering model, lifecycle, and communication between your MCP server and ChatGPT. A framework builds on top of that to give you the tooling you need to develop, test, and ship apps.

The comparison to React and Next.js still holds. React is the rendering library. Next.js gives you routing, server-side rendering, a dev server, and deployment. You can build a React app without Next.js, but most teams use a framework because it handles the infrastructure so you can focus on product.

A ChatGPT App framework does the same thing for ChatGPT Apps.

The Pain of Building Without a Framework

If you’ve tried building a ChatGPT App from the OpenAI Apps SDK alone, you’ve hit these problems:

No local testing. The only way to see your app render is to connect it to the real ChatGPT, which requires a paid ChatGPT Plus or Team subscription with developer mode. Every change means tunneling your local server, refreshing ChatGPT, and waiting for the round trip. That’s a 4-click manual refresh on every code change.

No component library. You build your own UI components from scratch, cards, layouts, forms, and get them looking right inside the conversation. You’re solving the same styling problems every time, including dark mode, display modes, and host-specific CSS variables.

No project structure. Every project starts from zero. There’s no standard way to organize your resources, tools, and configuration. You’re making structural decisions before you’ve written a line of product code.

No testing story. You can’t run automated tests against the ChatGPT interface. There’s no way to verify your app renders correctly in CI. Manual testing through the ChatGPT UI is the only option, and it burns host credits on every attempt.

No multi-host portability. If you want your app to also work in Claude, VS Code Copilot, or Goose, you’re managing host-specific protocol differences yourself. MCP Apps are portable in theory, but host-specific extensions (ChatGPT display modes, Claude data access patterns) require real abstraction work.

No deployment pipeline. Getting your app from local development to production means manually configuring an MCP server, setting up hosting, and wiring everything together.

What a ChatGPT App Framework Gives You

sunpeak maps each of those pain points to a concrete solution:

Local runtime inspector. Run pnpm dev and open localhost:3000. You get a full ChatGPT and Claude inspector that renders your app exactly as each host would, with no paid account, no tunneling, and no round trips. It supports all display modes, light and dark themes, and tool invocations.

pnpm dev
# Inspector running at http://localhost:3000
# MCP server running at http://localhost:8000

UI components. sunpeak includes production-ready components, cards, carousels, forms, and layouts, that follow host design conventions and handle cross-host styling automatically.

CLI scaffolding. Run npx sunpeak new and get a working project with dependencies installed, configuration set up, and a starter app ready to modify.

npx sunpeak new

Testing framework. sunpeak’s built-in testing framework runs against the inspector with no paid host accounts. It supports unit tests, Playwright-based e2e tests, visual regression tests with screenshot comparison, mocking and stubbing, and multi-model LLM evals. Everything runs locally and in CI/CD.

pnpm test              # Unit + e2e tests
pnpm test:visual       # Visual regression tests
pnpm test:eval         # Multi-model LLM evals

Multi-host portability. Write one codebase and it works across ChatGPT, Claude, VS Code, and Goose. Core APIs work everywhere. Host-specific features are opt-in via subpath imports (sunpeak/chatgpt, sunpeak/claude). You can build one MCP App for ChatGPT and Claude without maintaining separate codebases.

Build and deploy. pnpm build creates production bundles and pnpm start runs a zero-config production server. You can also mount on Express, Cloudflare Workers, or any Node-compatible runtime.

Building With vs. Without a Framework

Here’s the developer workflow side by side:

Without a framework:

# Set up a new project
mkdir sunpeak-app && cd sunpeak-app
pnpm init
pnpm add @modelcontextprotocol/sdk express
# Manually configure MCP server...
# Manually set up React rendering...
# Manually build UI components...
# Set up ngrok tunnel to test in ChatGPT...
# Hope it renders correctly...
# Repeat for Claude, VS Code, Goose...

With sunpeak:

# Set up a new project
npx sunpeak new
cd sunpeak-app

# Develop with instant feedback
pnpm dev
# Open localhost:3000 — done.

# Test across hosts automatically
pnpm test:visual

The difference goes beyond fewer commands. It’s fewer decisions, fewer things to debug, and fewer things that can go wrong before you’ve started building your product.

When You Don’t Need a Framework

A framework isn’t always the right choice.

If you’re building a server-side-only MCP tool that doesn’t render any UI in ChatGPT, you don’t need sunpeak. A plain MCP server with a few tool handlers is straightforward to set up with the @modelcontextprotocol/sdk package.

If you’re writing a one-off script or experimenting with the protocol, going framework-free is fine.

But the moment you’re building a UI that renders inside ChatGPT or Claude, especially one you plan to maintain, test, and ship to users, a framework pays for itself immediately. The MCP Apps protocol is still evolving, and host-specific extensions keep changing. A framework that stays current with those changes saves real maintenance time.

Get Started

sunpeak is open-source (MIT) and works locally and in CI/CD with no paid AI host accounts. Install it, scaffold a project, and start building.

npx sunpeak new
cd sunpeak-app && pnpm dev

See the quickstart guide to build your first ChatGPT App, or explore the MCP App framework page for the full feature set.