Workshop: Debug with Appium MCP

Watch this workshop, where 𝐒𝐫𝐢𝐧𝐢𝐯𝐚𝐬𝐚𝐧 𝐒𝐞𝐤𝐚𝐫, Director of Engineering at TestMu AI, and 𝐒𝐚𝐢 𝐊𝐫𝐢𝐬𝐡𝐧𝐚, Director of Engineering at TestMu AI, show how the Model Context Protocol turns Appium from a traditional automation framework into an AI-native testing platform that understands natural language and interacts intelligently with mobile apps. Learn how mcp-appium bridges AI assistants like Claude and Appium's mobile automation capabilities. Srinivasan and Sai also demonstrate what makes the approach different: as you explore your app through natural language commands, mcp-appium generates production-ready test code, turning exploratory sessions into maintainable automation suites - including how mcp-appium and mcp-webdriveragent work together to solve iOS real-device testing. 𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬: 0:00 Workshop Opens - Debugging With Appium MCP 1:12 Speaker Intros - Srinivasan Sekar and Sai Krishna, Directors of Engineering, TestMu AI 2:42 What MCP Actually Is: One Protocol, Tools, Resources, Prompts 4:18 The Client-Server Model and Dynamic Tool Loading 5:50 stdio vs Streamable HTTP, and Authenticated MCP Servers 6:53 Why Appium Built Its Own Official MCP Server 7:56 The Documentation Problem: Scattered Across 100+ Repos and Versions 8:58 Bringing the Docs Inside the MCP Server 9:28 Building and Signing WebDriverAgent - the Perennial iOS Pain 10:00 Generating Ranked Locator Strategies Instead of Hand-Hunting 11:02 Gestures: From Touch Actions to Actions API, Written for You 12:02 No Standalone Server Needed - Drivers Bundled In 13:39 Configuring the MCP in VS Code and Claude Code 15:10 Why Appium MCP Docs Is a Separate Opt-In Package 17:45 From 80 Tools Down to 33 - Why Tool Sprawl Is an Anti-Pattern 18:47 Designing Tools Around User Intent, Not API Endpoints 19:48 Building Your Own Appium MCP Plugin 21:52 Touring the Tools: Device Setup, WDA Prep, Inspection 23:25 Live Demo: Opening the Settings App by Natural Language 24:59 Watching It Select a Device and Create a Session 25:30 The App Lifecycle Tool as an Intent-Based Action 26:31 Q&A: Authentication and Sensitive Data With Local Models 28:05 Q&A: How MCP Changes Mobile Automation Architecture 29:11 Token Consumption: 13-14K vs 2K With an Optimised Agent 30:42 The Sample WebdriverIO Project 32:17 Live Demo: Generating Locators Across Three Screens 34:57 Ranked Locators and Best vs Alternate Strategies 36:30 Generating the Test for the Whole Journey 36:30 Q&A: Appium Plus MCP as a Foundation for Autonomous Mobile QA 38:34 Why Vision Models Matter When the Accessibility Tree Is Weak 39:35 Keeping Context Grounded and Avoiding Drift at Scale 41:11 Reviewing the Generated Locator Files and Ranking Logic 43:47 Breaking a Locator Deliberately and Watching the Test Fail 45:51 Q&A: Streaming Live Screenshots to a Vision Model 46:52 The MJPEG Server Trick: 20ms Instead of 250ms Screenshots 47:23 Q&A: Should Agents Auto-Heal Broken Locators in Real Time? 49:28 The Agent Fixes the Locator and the Missing Navigation 50:28 Asking the Docs Tool How to Run Parallel Sessions 51:29 RAG Over the Appium Docs: 25 Relevant Chunks Retrieved 53:02 The Answer Including WebdriverIO Config and Unique Ports 53:33 Q&A: Gesture Synchronisation Under Network and Rendering Lag 55:08 Optimising by Fetching Only Interactable Elements 57:14 Q&A: Pre-Building the WDA IPA as a One-Time CI Step 59:51 Q&A: The Real Latency Bottleneck in Live Debugging 1:01:25 Q&A: Certificate Expiry and Re-Signing WebDriverAgent 1:02:58 Q&A: Which Local Llama Models Actually Work 1:04:33 Why They Built AppClaw on Top of Appium MCP 1:05:34 Reducing Tokens: Screenshot Quality and App Guides 1:06:35 App Guides: Teaching the Agent Where Hidden Features Live 1:08:09 Context Compaction and Episodic Memory 1:09:13 Negative Caching So the Agent Doesn't Repeat a Failure 1:10:13 Proximity Selectors, Vision Plus DOM, and the AppClaw Runner 1:11:45 Comparing Token Cost: Claude Code vs AppClaw on the Same Goal 1:12:16 Playground Demo: Natural Language Commands and Test Export 1:15:21 Q&A: Reading Crash Logs From Physical Devices 1:16:25 Q&A: When an Agent Has Earned Enough Trust to Act 1:18:29 Q&A: Exporting a Discovered Test as a Deterministic Test 1:19:33 Q&A: Flutter, Virtualised Lists and Infinite Scroll Register for TestMuConf 2027: https://www.testmuai.com/testmuconf-2027/?utm_source=youtube&utm_medium=organic&utm_term=&utm_campaign=debug_with_appium_mcp #TestMuConf #TestMuAI #Appium #MCP #MobileTesting #TestAutomation #AITesting #Workshop