sunpeak is MCP-App-First (April 2026)

MCP Apps run in ChatGPT, Claude, VS Code, Goose, and JetBrains IDEs. When Claude announced MCP App support in January 2026 and ChatGPT followed in February, the standard went from spec to production in weeks. Since then, the ecosystem has grown fast: launch partners like Figma, Canva, Slack, Asana, and Salesforce have live MCP Apps in the directory, and the ext-apps spec is actively maintained under the Linux Foundation.

TL;DR: sunpeak is an MCP App framework built around the MCP App standard. Core APIs work across every host. ChatGPT and Claude-specific features are layered on top as optional imports. Write your app once, test it locally with a built-in inspector, and deploy it everywhere.

Why MCP-App-First

When ChatGPT Apps launched in October 2025, they had their own proprietary SDK. Building for ChatGPT meant building only for ChatGPT.

That changed when OpenAI contributed to MCP Apps as the open standard. The rendering model, the iframe sandbox, the UI resource system, all of it became portable. By early 2026, the major hosts shipped real implementations.

sunpeak followed the same trajectory. We started as a ChatGPT App framework because ChatGPT was the only major host supporting embedded UIs. Now we’re an MCP App framework, because the standard is real and the host list keeps growing.

What that means in practice: sunpeak’s core APIs target the MCP App interface, not any single host. At the same time, sunpeak layers in host-specific functionality that developers need to seamlessly support differentiated platforms (think React Native). You get portability by default and platform power when you need it.

How It Works

sunpeak separates standard MCP App APIs from host-specific ones at the import level.

Core APIs come from the top-level sunpeak import. These work in every host (see the React hooks reference for the full API):

import { useToolData, useHostContext, useDisplayMode, AppProvider } from 'sunpeak';
import type { ResourceConfig } from 'sunpeak';

export const resource: ResourceConfig = {
  description: 'Show analytics dashboard',
};

export function DashboardResource() {
  const { output } = useToolData();
  const context = useHostContext();
  const displayMode = useDisplayMode();

  return <div>{/* Your UI runs in ChatGPT, Claude, VS Code, Goose */}</div>;
}

Inspector and testing APIs come from dedicated subpath imports. sunpeak/inspector gives you the Inspector component for local development, and the testing framework provides unit tests, e2e tests, visual regression tests, and multi-model LLM evals:

pnpm dev           # Inspector at localhost:3000
pnpm test          # Unit + e2e tests
pnpm test:visual   # Visual regression across hosts/themes
pnpm test:eval     # Multi-model evals (GPT-4o, Claude, Gemini)

The inspector, testing utilities, and dev tools live in dedicated subpaths. They’re first-class, not afterthoughts, but they don’t pollute your app code. Your resource components stay portable.

The Host List

Here’s where MCP Apps run today:

• ChatGPT — OpenAI contributed elements of the original ChatGPT Apps protocol to MCP and supports the open standard. Inline, picture-in-picture, and fullscreen display modes are available.
• Claude — Anthropic’s web and desktop clients render MCP Apps natively via Claude Connectors. First-party integrations from Figma, Canva, Asana, Slack, and others are live in the Connectors Directory.
• VS Code — Microsoft’s editor renders MCP Apps in the Copilot chat sidebar.
• Goose — Block’s open-source AI agent supports MCP Apps.
• JetBrains IDEs — JetBrains is evaluating MCP App support across their IDE lineup.

The MCP App standard is under the Linux Foundation, and the spec is actively developed at modelcontextprotocol/ext-apps. The ui:// URI scheme lets tools declare HTML templates that hosts prefetch and render in sandboxed iframes, with JSON-RPC 2.0 over postMessage handling communication. Any compliant host can render them.

Host-Specific Features Are First-Class

MCP-App-first doesn’t mean lowest-common-denominator. ChatGPT has features that Claude doesn’t, and vice versa. sunpeak treats these as first-class extensions, not hacks.

For ChatGPT, that means the ChatGPT inspector for local development, display mode control, and ChatGPT-specific runtime APIs. For Claude, it means Connector data access patterns, OAuth authentication, and interactive Connector support. These are maintained alongside the core framework, tested in CI, and documented.

The pattern scales: core stays portable, extensions stay organized. Import from sunpeak for the standard, import from sunpeak/chatgpt or sunpeak/claude for the host-specific parts.

Testing Across Hosts

One of the practical benefits of MCP-App-first architecture is testing. Because sunpeak’s inspector replicates the MCP App runtime that all hosts implement, you can run your full test suite locally without paid accounts or host credits.

sunpeak’s testing framework includes unit tests, Playwright-based e2e tests, snapshot tests, visual regression tests with screenshot comparison across hosts and themes, and multi-model LLM evals that verify tool calling across GPT-4o, Claude, and Gemini. All of it runs locally and in CI/CD.

pnpm test          # Unit + e2e
pnpm test:visual   # Screenshot comparison across ChatGPT/Claude themes
pnpm test:eval     # Multi-model LLM evals

Get Started

sunpeak is open-source (MIT) and works locally and in CI/CD with no paid AI host accounts. Scaffold a project, start the inspector, and build your first MCP App.

npx sunpeak new
cd sunpeak-app && pnpm dev

See the quickstart guide to get started, or explore the MCP App framework page for the full feature set.