MCP App Testing Framework

Test CLI

How It Works

import { test, expect } from 'sunpeak/test';

test('albums render in light mode', async ({ inspector }) => {
  const result = await inspector.renderTool('show-albums', undefined, { theme: 'light' });
  const app = result.app();
  await expect(app.locator('button:has-text("Summer Slice")')).toBeVisible();
});

test('albums render in fullscreen', async ({ inspector }) => {
  const result = await inspector.renderTool('show-albums', undefined, { displayMode: 'fullscreen' });
  const app = result.app();
  await expect(app.locator('button:has-text("Summer Slice")')).toBeVisible();
  // Compare against saved baseline (only runs with --visual flag)
  await result.screenshot('albums-fullscreen');
});

What You Can Test

• Multi-Host Rendering

  Tests run against both ChatGPT and Claude runtimes automatically via Playwright projects. One test file covers both hosts.

• Themes & Display Modes

  Test light/dark themes and inline/fullscreen/pip display modes. Use setTheme() and setDisplayMode() or pass options to callTool().

• Visual Regression

  Capture screenshots with result.screenshot() and compare against baselines. Configure thresholds and max diff pixel ratios in defineConfig().

• Backend Tool Mocking

  Simulation files can mock callServerTool responses with simple or conditional matching. Test interactive flows without a real backend.

• MCP-Specific Assertions

  Custom matchers: toHaveTextContent(), toHaveStructuredContent(), toBeError() alongside standard Playwright locators.

• Multi-Model Evals

  Send prompts to GPT-4o, Claude, Gemini, and other models. Assert each model calls the right tools with the right arguments. Each eval runs N times per model to measure reliability.

• Any MCP Server, Any Language

  Use npx sunpeak test init with any MCP server. Configure the server via HTTP URL or startup command. Python, Go, TypeScript, Rust, anything.

Who It's For