Download to read offline
Ready Layer One: Intro to the Model Context Protocol
- 1. Intro to theModel Context Protocol READY LAYER ONE MATT MCKENNA #DCNY25 Matt McKenna #dcny25
- 2.
- 3.
- 4. WHATWE’LLCOVERTODAY High level viewof LLMs and AI Agents What is the Model Context Protocol Android Device control with AI
- 5. TAKEAWAYS Demystify the ModelContext Protocol Enable you all to build your own MCP Servers!
- 6.
- 7. - Sequence predictionat a massive scale WHATISALARGELANGUAGEMODEL(LLM)
- 8. - Sequence predictionat a massive scale - Trained on broad data, fi ne tuned on task speci fi c patterns WHATISALARGELANGUAGEMODEL(LLM)
- 9. - Sequence predictionat a massive scale - Trained on broad data, fi ne tuned on task speci fi c patterns - Interfaces WHATISALARGELANGUAGEMODEL(LLM)
- 10. - Sequence predictionat a massive scale - Trained on broad data, fi ne tuned on task speci fi c patterns - Interfaces - Chat - Code Completion - APIs WHATISALARGELANGUAGEMODEL(LLM)
- 11.
- 12. - You givethe model some text (this is a prompt) HOWDOESANLLMWORK?
- 13. - You givethe model some text (this is a prompt) - It tries to predict what comes next a piece at a time HOWDOESANLLMWORK?
- 14. - You givethe model some text (this is a prompt) - It tries to predict what comes next a piece at a time - It doesn’t “think” or “know”, but follows patterns HOWDOESANLLMWORK?
- 15.
- 16.
- 17. The android developer predictnext invalidates
- 18. The android developer predictnext invalidates
- 19. The android developer predictnext invalidates caches
- 20. The android developer predictnext caches invalidates
- 21. The android developer predictnext caches invalidates and
- 22. The android developer predictnext caches and invalidates
- 23. The android developer predictnext restarts caches and invalidates
- 24. The android developer predictnext invalidates caches and restarts
- 25.
- 26. - Let’s dosomething with the Android Debug Bridge! (ADB) LLMSINACTION
- 27. - Let’s dosomething with the Android Debug Bridge! (ADB) - Hey can you write some adb commands for me to launch the Bluesky android app and input text saying “hey from droidcon new york!” and then post it? LLMSINACTION
- 28.
- 29.
- 30.
- 31.
- 32.
- 33.
- 34.
- 35.
- 36.
- 37.
- 38.
- 39.
- 40.
- 41.
- 42.
- 43.
- 44. WHATISANAIAGENT? A system thatuses a language model to complete tasks autonomously or semi-autonomously
- 45.
- 46. - A languagemodel WHATISANAIAGENT?
- 47. - A languagemodel - Tools it can call WHATISANAIAGENT?
- 48. - A languagemodel - Tools it can call - A planner to decide on and sequence actions WHATISANAIAGENT?
- 60.
- 61.
- 62.
- 63.
- 64. - Tools arethe capabilities of the agent AIAGENTSNEEDTOOLS
- 65. - Tools arethe capabilities of the agent - Structured input / output AIAGENTSNEEDTOOLS
- 66. - Tools arethe capabilities of the agent - Structured input / output - Need to describe - What’s available - When to use - How to use AIAGENTSNEEDTOOLS
- 67.
- 68.
- 69.
- 70. - A JSONbased protocol for tool description and invocation THEMODELCONTEXTPROTOCOL
- 71. - A JSONbased protocol for tool description and invocation - Share contextual information with LLMs THEMODELCONTEXTPROTOCOL
- 72. - A JSONbased protocol for tool description and invocation - Share contextual information with LLMs - Expose tools and capabilities THEMODELCONTEXTPROTOCOL
- 73. - A JSONbased protocol for tool description and invocation - Share contextual information with LLMs - Expose tools and capabilities - Build composable integrations and work fl ows THEMODELCONTEXTPROTOCOL
- 74. - A JSONbased protocol for tool description and invocation - Share contextual information with LLMs - Expose tools and capabilities - Build composable integrations and work fl ows - Enables tools to describe themselves THEMODELCONTEXTPROTOCOL
- 75. - A JSONbased protocol for tool description and invocation - Share contextual information with LLMs - Expose tools and capabilities - Build composable integrations and work fl ows - Enables tools to describe themselves THEMODELCONTEXTPROTOCOL
- 76.
- 77. - Resources - File-likedata that can be read by clients MCPFEATURES
- 78. - Resources - File-likedata that can be read by clients - Prompts - Pre-written templates that help Users accomplish speci fi c tasks MCPFEATURES
- 79. - Resources - File-likedata that can be read by clients - Prompts - Pre-written templates that help Users accomplish speci fi c tasks - Tools - Functions that can be called by the LLM MCPFEATURES
- 80.
- 81. - Easy tobuild! MCPSERVERS
- 82. - Easy tobuild! - Expose Resource, Tools, and Prompts via primitives MCPSERVERS
- 83. - Easy tobuild! - Expose Resource, Tools, and Prompts via primitives - Can work with local processes and/or remote services MCPSERVERS
- 84. ARCHITECTURE AI AGENT /HOST ADB Weather Calendar Remote Data Source Local Data Source Local Data Source
- 85. ARCHITECTURE AI AGENT /HOST ADB MCP Weather MCP Calendar MCP Remote Data Source Local Data Source Local Data Source Mode l Cont ext Proto col
- 86.
- 87.
- 88.
- 89. fun adb(vararg args:String): String { }
- 90. fun adb(vararg args:String): String { val command = listOf("adb") + args }
- 91. fun adb(vararg args:String): String { val command = listOf("adb") + args val process = Runtime.getRuntime().exec(command.toTypedArray()) }
- 92. fun adb(vararg args:String): String { val command = listOf("adb") + args val process = Runtime.getRuntime().exec(command.toTypedArray()) return process.inputStream.bu ff eredReader().readText() }
- 93.
- 94.
- 95.
- 96.
- 97.
- 98. fun main(args: Array<String>){ val command = args. fi rstOrNull() ?: "--sse-server-ktor" val port = args.getOrNull(1)?.toIntOrNull() ?: 3001 when (command) { "--stdio" -> runMcpServerUsingStdio() "--sse-server-ktor" -> runSseMcpServerUsingKtorPlugin(port) "--sse-server" -> runSseMcpServerWithPlainCon fi guration(port) else -> { System.err.println("Unknown command: $command") } } }
- 99. fun runSseMcpServerUsingKtorPlugin(port: Int):Unit = runBlocking { println("Starting sse server on port $port") println("Use inspector to connect to the http://localhost:$port/sse") embeddedServer(CIO, host = "0.0.0.0", port = port) { MCP { return@MCP con fi gureServer() } }.start(wait = true) }
- 100. fun con fi gureServer(): Server{ val def = CompletableDeferred<Unit>() val server = Server( Implementation( name = "mcp-server-android", version = "0.1.0" ), ServerOptions( capabilities = ServerCapabilities( prompts = ServerCapabilities.Prompts(listChanged = true), resources = ServerCapabilities.Resources(subscribe = true, listChanged = true), tools = ServerCapabilities.Tools(listChanged = true), ) ), onCloseCallback = { def.complete(Unit) } ) //...
- 101. fun con fi gureServer(): Server{ //... // Create AdbManager instance val deviceManager = AdbManager() //...
- 102. fun con fi gureServer(): Server{ //... // Create AdbManager instance val deviceManager = AdbManager() // Create AndroidToolsRegistry instance and register all tools val toolsRegistry = AndroidToolsRegistry(deviceManager) toolsRegistry.registerTools(server) //...
- 103. fun con fi gureServer(): Server{ //... // Add Android device control prompt server.addPrompt( name = "Android Device Control", description = "Control and interact with Android devices through MCP.", arguments = listOf() ) { _ -> // prompt of all tools } return server }
- 104.
- 105.
- 106. - Start withwork fl ows, not raw endpoints DESIGNTOOLSFROMWORKFLOWS
- 107. - Start withwork fl ows, not raw endpoints - Combine steps into a single high level Tool DESIGNTOOLSFROMWORKFLOWS
- 108. - Tool namesand params ARE prompts PROMPTFRIENDLYTOOLDESIGN
- 109. - Tool namesand params ARE prompts - Better Tool descriptions mean better tool use PROMPTFRIENDLYTOOLDESIGN
- 110. - LLMs contextis limited PROTECTTHECONTEXTWINDOW
- 111. - LLMs contextis limited - Running out of context window is costly PROTECTTHECONTEXTWINDOW
- 112. - LLMs contextis limited - Running out of context window is costly - MCPs can help PROTECTTHECONTEXTWINDOW
- 113.
- 114. - Join thecommunity on Bluesky BLUESKY go.bsky.app/Etd57t6
- 115. - Kotlin Multiplatform implementationof MCP MCPKOTLINSDK MODEL CONTEXT PROTOCOL github.com/modelcontextprotocol/kotlin-sdk
- 116. - A local,open source AI agent, automating engineering tasks seamlessly. GOOSE AI AGENT BY BLOCK block.github.io/goose/
- 117. - Best practicesand learnings for designing MCP Servers DESIGNINGMCPSERVERS BLOCK’S PLAYBOOK FOR engineering.block.xyz/blog/blocks-playbook-for-designing-mcp-servers
- 118.
- 119. - Enable Androidapps to provide MCP server access to AI tools in development environments ANDROID-MCP-SDK JASON PEARSON’S github.com/kaeawc/android-mcp-sdk
- 120. - Introducing Trailblaze,a natural language AI driven test framework. AIMOBILETESTING - BLAZINGTHETRAILATBLOCK SAM EDWARDS + BRIAN GARDNER STAGE 2 Catch the VOD
- 121.
- 122.