sunpeak: Instructions file for Codex

sunpeak

Note that "sunpeak", except where required in URLs or code, is always lowercase.

sunpeak is a framework for building MCP Apps with interactive UIs that run inside AI chat hosts (ChatGPT, Claude, and future major hosts). Built on top of the MCP Apps SDK (@modelcontextprotocol/ext-apps).

ChatGPT apps are now submitted and published as plugins. The app portion is still based on MCP Apps or the OpenAI Apps SDK, so sunpeak's protocol and runtime architecture do not change. Adding a custom app requires ChatGPT Developer mode, enabled from the bottom-left user menu under Settings > Security and login > Developer mode. Local developer-mode apps are then created from the ChatGPT homepage under Plugins > +, while public apps are submitted as MCP-backed plugins through the OpenAI Platform.

The value proposition of the sunpeak framework is to help developers and their agents:

1. Test MCP Apps locally and automatically (in CI/CD) using a replica of the ChatGPT and Claude runtimes.
   1. Save time manually testing all possible host, server, app, ui, and backend states.
   2. Protect developers from 4-click manual refreshes on every code change in each host.
   3. Cancel all the $20 per person per host per month testing accounts.
   4. Avoid burning host credits on every test and code change.
2. Verify MCP tools work across multiple LLM models via the eval framework. Evals connect to the MCP server, discover tools, and send prompts to multiple models (GPT-4o, GPT-4o-mini, o4-mini, Claude Sonnet, Gemini 2.0 Flash) via the Vercel AI SDK. Cases can include App Context for follow-up turns that depend on model-visible UI state. Each case runs N times per model and reports pass/fail counts, so developers can confirm that tool descriptions, schemas, and App Context work reliably on smaller and cheaper models. Opt-in via sunpeak test --eval because evals cost money.
3. Automate the real-host testing loop with live tests (sunpeak test --live): scripts that drive a real browser into ChatGPT (and other hosts as they're supported), send prompts that trigger MCP tool calls against the developer's server, and assert against the actually-rendered app via Playwright. Live tests catch what the inspector can't (real MCP connection behavior, real LLM tool invocation, host-specific iframe rendering, production resource loading) and replace the manual prompt-and-click loop. Opt-in because they hit real accounts.
4. Build multi-platform MCP Apps in a structured way that's easy to understand and get started.
5. Test their MCPs in ChatGPT with HMR and Claude with automatic rebuilds and refresh notifications.
6. Embed the Inspector as a React component inside any third-party React app to render arbitrary MCP Apps. The <Inspector app={...} onCallTool={...} /> shape exposes the same double-iframe runtime, host shells, and conversation chrome as the CLI inspector, but without any sunpeak-owned servers or /__sunpeak/* runtime dependencies. Tool calls flow through a callback the embedder owns; resource HTML is passed as a string. Embedders host the static sandbox proxy (dist/sandbox-proxy.html) on a separate origin from their own app — or fall back to the same-origin srcdoc variant for zero deployment. This is a first-class use case; treat it as load-bearing when designing changes.