Flutter Skill

flutter-skill

Give any AI agent eyes and hands inside any running app.

10 platforms. Zero test code. One MCP server.

Demo • Quick Start • AI Platforms • Platforms • vs Others • Docs

🚀 Zero config. Zero test code. Just talk to your AI.

If this saves you time, please consider starring the repo ⭐ — it helps others find it!

---

30-Second Demo

https://github.com/user-attachments/assets/d4617c73-043f-424c-9a9a-1a61d4c2d3c6

One prompt. 28 AI-driven actions. Zero test code. The AI explores a TikTok clone, navigates tabs, scrolls feeds, tests search, fills forms — all autonomously.

---

Why This Exists

Writing E2E tests is painful. Maintaining them is worse. flutter-skill takes a different approach:

• 🔌 Connects any AI agent (Claude, Cursor, Windsurf, Copilot, OpenClaw) directly to your running app via MCP
• 👀 The agent sees your screen — taps buttons, types text, scrolls, navigates — like a human tester who never sleeps
• ✅ Zero test code — no Page Objects, no XPath, no brittle selectors. Just plain English
• ⚡ Zero config — 2 lines of code, works on all 10 platforms

You: "Test the checkout flow with an empty cart, then add 3 items and complete purchase"

Your AI agent handles the rest — screenshots, taps, text entry, assertions, navigation. No Page Objects. No XPath. No brittle selectors. Just plain English.

---

Quick Start

1. Install (30 seconds)

npm install -g flutter-skill

2. Add to your AI (copy-paste into MCP config)

{

"mcpServers": { "flutter-skill": { "command": "flutter-skill", "args": ["server"] } } }

Works with Claude Desktop, Cursor, Windsurf, Copilot, Cline, OpenClaw — any MCP-compatible agent.

3. Add to your app (2 lines for Flutter)

import 'package:flutter_skill/flutter_skill.dart';

void main() { if (kDebugMode) FlutterSkillBinding.ensureInitialized(); runApp(MyApp()); }

4. Test — just talk to your AI:

*"Launch my app, explore every screen, and report any bugs"*

That's it. Zero configuration. Zero test code. Works in under 60 seconds.

📦 More install methods (Homebrew, Scoop, Docker, IDE, Agent Skill)

---

Use with AI Platforms

MCP Server Mode (IDE Integration)

Works with any MCP-compatible AI tool. One config line:

{

"mcpServers": { "flutter-skill": { "command": "flutter-skill", "args": ["server"] } } }

HTTP Serve Mode (CLI & Automation)

For standalone browser automation, CI/CD pipelines, or remote access:

# Start server

flutter-skill serve https://your-app.com

Use CLI client commands

flutter-skill nav https://google.com flutter-skill snap # Accessibility tree (99% fewer tokens) flutter-skill screenshot /tmp/ss.jpg flutter-skill tap "Login" flutter-skill type "hello@example.com" flutter-skill eval "document.title" flutter-skill tools # List all available tools

Supports --port=N, --host=H flags and FS_PORT/FS_HOST env vars.

Two Modes Compared

Full CLI client reference: docs/CLI_CLIENT.md

---

10 Platforms, One Tool

Most testing tools work on 1-2 platforms. flutter-skill works on 10.

Total: 656/664 tests passing (98.8%) — each platform tested against a complex social media app with 50+ elements.

---

⚡ Performance

Real benchmarks from automated test runs against a complex social media app:

Token efficiency: snapshot() returns a structured element tree instead of an image — 87–99% fewer tokens than sending screenshots to your AI agent.

How fast is that? A tap takes 1–2 ms end-to-end. Browser automation tools like Playwright and Selenium typically take 50–100 ms for the same operation. That's 50–100× faster, because flutter-skill talks directly to the app runtime instead of going through WebDriver or CDP indirection.

Heavy DOM Sites (Real-World)

Tested 15 MCP tools against production websites — 75/75 passed, zero timeouts:

Total time includes page load. Tool execution is consistently sub-100ms even on heavy DOM sites.

---

Why Not Playwright / Appium / Detox?

| AI page understanding | ✅ AX Tree | ❌ Screenshots | ❌ | ❌ | | Boundary/security test | ✅ 13 payloads | ❌ | ❌ | ❌ | | Batch actions | ✅ 5+/call | 1/call | 1/call | 1/call |

flutter-skill is the only AI-native E2E testing tool that works across mobile, web, and desktop — with 7× more tools than the nearest competitor.

---

CLI Commands

# 🤖 AI autonomous exploration — finds bugs automatically

flutter-skill explore https://my-app.com --depth=3

🐒 Monkey/fuzz testing — random actions, crash detection

flutter-skill monkey https://my-app.com --actions=100 --seed=42

🚀 Parallel multi-platform testing

flutter-skill test --url https://my-app.com --platforms web,electron,android

🌐 Zero-config WebMCP server — any website becomes testable

flutter-skill serve https://my-app.com

---

🧠 AI-Native: 95% Fewer Tokens

Most AI testing tools send screenshots to the LLM — each one costs ~4,000 tokens.

flutter-skill uses Chrome's Accessibility Tree to give your AI a compact semantic summary of any page:

// page_summary → ~200 tokens (vs ~4,000 for a screenshot)

{ "title": "Shopping Cart", "nav": ["Home", "Products", "Cart", "Account"], "forms": [{"input:Coupon Code": "text"}], "buttons": ["Apply", "Checkout", "Continue Shopping"], "features": {"search": true, "pagination": true}, "links": 47, "inputs": 3 }

Then batch multiple actions in one call:

// explore_actions → 5 actions per call (vs 5 separate tool calls)

{"actions": [ {"type": "fill", "target": "input:Coupon Code", "value": "SAVE20"}, {"type": "tap", "target": "button:Apply"}, {"type": "tap", "target": "button:Checkout"}, {"type": "fill", "target": "input:Email", "value": "test@example.com"}, {"type": "tap", "target": "button:Continue"} ]}

Result: Your AI agent tests faster, costs less, and understands pages better than screenshot-based tools.

| Tokens per page | **~2