Workshop: Debug with Appium MCP

on YouTube ↗01h 23m

" summary

Appium is extended into an AI-native testing platform that understands natural language, uses mcp-appium to bridge Claude-style agents and Appium drivers, and generates production-ready WebdriverIO tests and ranked locator files on the fly. Workshop demos show the iOS pain points solved by mcp-webdriveragent (pre-build WDA IPA for CI), tool count reduced from ~80 to 33, an MJPEG server trick cutting screenshots to ~20ms, and token costs of 13–14K vs 2K tokens with optimised agents.

AI-written summary · about

" chapters (57)

1. 0:00Workshop Opens - Debugging With Appium MCP
2. 1:12Speaker Intros - Srinivasan Sekar and Sai Krishna, Directors of Engineering, TestMu AI
3. 2:42What MCP Actually Is: One Protocol, Tools, Resources, Prompts
4. 4:18The Client-Server Model and Dynamic Tool Loading
5. 5:50stdio vs Streamable HTTP, and Authenticated MCP Servers
6. 6:53Why Appium Built Its Own Official MCP Server
7. 7:56The Documentation Problem: Scattered Across 100+ Repos and Versions
8. 8:58Bringing the Docs Inside the MCP Server
9. 9:28Building and Signing WebDriverAgent - the Perennial iOS Pain
10. 10:00Generating Ranked Locator Strategies Instead of Hand-Hunting
11. 11:02Gestures: From Touch Actions to Actions API, Written for You
12. 12:02No Standalone Server Needed - Drivers Bundled In
13. 13:39Configuring the MCP in VS Code and Claude Code
14. 15:10Why Appium MCP Docs Is a Separate Opt-In Package
15. 17:45From 80 Tools Down to 33 - Why Tool Sprawl Is an Anti-Pattern
16. 18:47Designing Tools Around User Intent, Not API Endpoints
17. 19:48Building Your Own Appium MCP Plugin
18. 21:52Touring the Tools: Device Setup, WDA Prep, Inspection
19. 23:25Live Demo: Opening the Settings App by Natural Language
20. 24:59Watching It Select a Device and Create a Session
21. 25:30The App Lifecycle Tool as an Intent-Based Action
22. 26:31Q&A: Authentication and Sensitive Data With Local Models
23. 28:05Q&A: How MCP Changes Mobile Automation Architecture
24. 29:11Token Consumption: 13-14K vs 2K With an Optimised Agent
25. 30:42The Sample WebdriverIO Project
26. 32:17Live Demo: Generating Locators Across Three Screens
27. 34:57Ranked Locators and Best vs Alternate Strategies
28. 36:30Generating the Test for the Whole Journey
29. 38:34Why Vision Models Matter When the Accessibility Tree Is Weak
30. 39:35Keeping Context Grounded and Avoiding Drift at Scale
31. 41:11Reviewing the Generated Locator Files and Ranking Logic
32. 43:47Breaking a Locator Deliberately and Watching the Test Fail
33. 45:51Q&A: Streaming Live Screenshots to a Vision Model
34. 46:52The MJPEG Server Trick: 20ms Instead of 250ms Screenshots
35. 47:23Q&A: Should Agents Auto-Heal Broken Locators in Real Time?
36. 49:28The Agent Fixes the Locator and the Missing Navigation
37. 50:28Asking the Docs Tool How to Run Parallel Sessions
38. 51:29RAG Over the Appium Docs: 25 Relevant Chunks Retrieved
39. 53:02The Answer Including WebdriverIO Config and Unique Ports
40. 53:33Q&A: Gesture Synchronisation Under Network and Rendering Lag
41. 55:08Optimising by Fetching Only Interactable Elements
42. 57:14Q&A: Pre-Building the WDA IPA as a One-Time CI Step
43. 59:51Q&A: The Real Latency Bottleneck in Live Debugging
44. 1:01:25Q&A: Certificate Expiry and Re-Signing WebDriverAgent
45. 1:02:58Q&A: Which Local Llama Models Actually Work
46. 1:04:33Why They Built AppClaw on Top of Appium MCP
47. 1:05:34Reducing Tokens: Screenshot Quality and App Guides
48. 1:06:35App Guides: Teaching the Agent Where Hidden Features Live
49. 1:08:09Context Compaction and Episodic Memory
50. 1:09:13Negative Caching So the Agent Doesn't Repeat a Failure
51. 1:10:13Proximity Selectors, Vision Plus DOM, and the AppClaw Runner
52. 1:11:45Comparing Token Cost: Claude Code vs AppClaw on the Same Goal
53. 1:12:16Playground Demo: Natural Language Commands and Test Export
54. 1:15:21Q&A: Reading Crash Logs From Physical Devices
55. 1:16:25Q&A: When an Agent Has Earned Enough Trust to Act
56. 1:18:29Q&A: Exporting a Discovered Test as a Deterministic Test
57. 1:19:33Q&A: Flutter, Virtualised Lists and Infinite Scroll

" related talks

• 23mAppium: Under the Hood of WebDriverAgentMykola Mokhnach · Appium Conference 2021
• 45mAI-Powered Debugging & Browser Automation with Playwright MCPDebbie O'Brien · TestMu Conference 2025
• 48mSupercharged Testing: AI-Powered Workflows with Playwright + MCPDebbie O'Brien · NDC London 2026
• 57mAppium: Automation for AppsDan Cuellar · Mobius 2017 Piter
• 24mAutomated Mobile E2E TestingHe Ye · iOS Conf SG 2025
• 39mQ & A With The Appium Committee PanelDan Cuellar, Naresh Jain · AppiumConf 2019
• 52mBuilding an Advanced Automation Solution Based on AppiumAnton Sirota · Selenium Camp 2019
• 44mIntercepting Network traffic : Appium Intercept Plugin WorkshopAnil Patidar · Appium Conf 2024