4 Requests, 2 Approvals: How MCP Works, Message by Message

Five AI apps that each need the same ten tools means fifty custom integrations, and any of them can break when either side changes. MCP, the Model Context Protocol, turns those fifty into fifteen. Here's what an MCP server actually is, message by message, and where it breaks. You'll learn: → What MCP is, and why apps × tools becomes apps + tools → The parts: the host, one client per server, the server, and the model that never connects to it → Local servers over standard input and output vs remote servers over HTTP → Tools, resources and prompts: who chooses each one → What goes over the wire: JSON-RPC requests that carry their own version and capabilities since the 2026-07-28 revision → A worked example, request by request: filing a bug, a tool error the model fixes itself, and two approvals → Where it breaks: untrusted tool text and prompt injection, a server's own bugs (a June 2025 flaw Asana disclosed in its 10-Q), and tool definitions that fill the context → Where MCP runs today, and MCP gateways on AWS, Azure, Google Cloud and in open source Chapters: 0:00 Fifty integrations, or fifteen 0:52 One connector per app, per tool 1:52 A plug, with a brain on one side 2:51 Local and remote servers 3:45 What goes over the wire 4:54 Worked example: filing a bug 6:09 Where it breaks 7:29 Where MCP runs today 8:33 Recap Sources: the specification, docs and filings this video explains. The worked example's tool names, labels and issue number are made up and labelled as an example; its message shapes follow the specification. Key links: → Model Context Protocol, What is MCP?: https://modelcontextprotocol.io/docs/... → MCP specification, revision 2026-07-28: https://modelcontextprotocol.io/speci... → MCP specification, Tools: https://modelcontextprotocol.io/speci... → MCP docs, Architecture overview: https://modelcontextprotocol.io/docs/... → MCP blog, The 2026-07-28 Specification (July 28, 2026): https://blog.modelcontextprotocol.io/... → MCP docs, Security Best Practices: https://modelcontextprotocol.io/docs/... → MCP docs, Client Best Practices (progressive discovery): https://modelcontextprotocol.io/docs/... → Anthropic, Introducing the Model Context Protocol (November 25, 2024): https://www.anthropic.com/news/model-... → Anthropic, Donating the Model Context Protocol and establishing the Agentic AI Foundation (December 9, 2025): https://www.anthropic.com/news/donati... → Anthropic Engineering, Code execution with MCP (November 4, 2025): https://www.anthropic.com/engineering... → Asana, Form 10-Q for the quarter ended July 31, 2025 (filed September 3, 2025): https://www.sec.gov/Archives/edgar/da... → Amazon Bedrock AgentCore Gateway: https://docs.aws.amazon.com/bedrock-a... → Azure API Management, MCP servers overview: https://learn.microsoft.com/en-us/azu... → Google Cloud, MCP support for Apigee (December 10, 2025): https://cloud.google.com/blog/product... → agentgateway docs: https://agentgateway.dev/docs/standal... Watch next: An AI Escaped Its Test Sandbox: The Hugging Face Incident Explained:    • An AI Escaped Its Test Sandbox: The Huggin...   If this helped, subscribe. There's a new breakdown every week. Model Context Protocol and MCP are names of an open-source project hosted by the Agentic AI Foundation, part of the Linux Foundation; neither the project nor its foundations sponsor or endorse this video. agentgateway is a project of the Agentic AI Foundation; neither the project nor the foundation sponsors or endorses this video. Anthropic and Asana are named only as the sources of published facts; neither sponsors nor endorses this video. ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code are named as products that adopted MCP; their owners do not sponsor or endorse this video. AWS, Microsoft Azure, Google Cloud, Apigee and their product names and logos belong to their owners, none of whom sponsors or endorses this video. #mcp #ai #computerscience