# How to vibe code responsibly (with a little help from MCPs) Source: https://www.youtube.com/watch?v=PvdptUZ3XeU Recap page: https://rapidrecap.app/video/PvdptUZ3XeU Generated: 2025-11-29T15:33:06.091+00:00 --- ## Quick Overview Rizel Scarlett advocates for "vibe coding responsibly" by choosing competent AI Large Language Models (LLMs) with tool-calling capabilities, using a Model Context Protocol (MCP) for structured context management, and leveraging the GitHub MCP Server to integrate agent actions like creating issues and fetching calendar data directly into development workflows. **Key Points:** - Rizel Scarlett introduced the concept of "vibe coding responsibly" to counteract the dangers of unchecked AI code generation, which can lead to insecure code, poor scalability, and lost work. - Responsible AI coding requires developers to maintain control, context, and power over the AI, rather than blindly accepting generated code. - Key factors when choosing an LLM for coding agents include Tool Calling capability, Context Window Size (memory), and Coding performance on relevant tasks. - The Model Context Protocol (MCP) is an open standard enabling any supporting AI agent to connect to and grab data from any tool using natural language prompts instead of specific syntax. - The GitHub MCP Server allows agents to interact with GitHub, demonstrated by successfully creating a GitHub issue titled "Speak at GitHub Universe" and assigning it to the speaker's handle (@blackgirlbytes). - The MCP also enables agents to interact with other tools, such as fetching a team's project calendar events, which is visualized as a literal calendar within the chat interface. - The presentation emphasized that developers must remain engaged, reviewing changes and reasoning about the AI's output rather than forgetting the code exists (Andrej Karpathy's definition of vibe coding). ![Screenshot at 00:07: The opening slide for the presentation titled "How to Vibe Code Responsibly With a little help from MCPs," introducing the speaker Rizel Scarlett, Tech Lead / Block, and the conference context \(Q25\).](https://ss.rapidrecap.app/screens/PvdptUZ3XeU/00-00-07.png) **Context:** The presentation, delivered by Rizel Scarlett, Tech Lead of Open Source Developer Relations at Block, focuses on integrating Large Language Models (LLMs) and AI agents into the software development lifecycle responsibly. Scarlett uses a 'Never Have I Ever' game and analogies like the 'elephant in the room' (unmanaged AI code) to set the stage for why developers need control and structure when using AI coding assistants, leading into her proposed solutions involving the Model Context Protocol (MCP) and the GitHub MCP Server. ## Detailed Analysis Rizel Scarlett argues that while AI coding assistants offer speed and empower non-engineers, relying solely on 'vibe coding' (writing code based on feeling without deep understanding) is dangerously irresponsible, leading to insecure code, scalability issues, and potential data loss exemplified by a Twitter user losing two weeks of work. To combat this, she proposes three principles for responsible AI usage: Power (choosing the right LLM), Context (feeding the right information), and Control. Power involves selecting LLMs that support features like Tool Calling and have adequate Context Window Size. Context is managed via an Agent Client Protocol (ACP), which allows agents to interact with tools using natural language, unlike older, syntax-heavy methods. She demonstrated this using the GitHub MCP Server, which allows an agent to interact with GitHub (creating an issue) and other integrated tools (like fetching a team calendar), providing a rich, interactive interface rather than just text output. The core message is that developers must remain in control, critically reviewing AI output and understanding the underlying logic, using these protocols to maintain oversight. ### Introduction & The Problem - Participants played 'Never Have I Ever' covering actions like asking AI instead of Google or merging AI code they didn't understand - This practice leads to insecure code, poor scalability, and loss of code. ### Power - Choosing Your LLM Wisely: Developers must select LLMs based on Tool Calling, Context Window Size (memory), and Coding performance, noting that some models are better at reasoning through complex code than others. ### Context - Feed your AI Agent the Right Information: Context is crucial; without it, AI tools may hallucinate or produce functionally correct but logically flawed code - The speaker used an analogy of choosing a handbag from a large collection to illustrate decision paralysis without context. ### Control - Agent Client Protocol (ACP): This open standard allows developers to give agents direct control over tools (like GitHub or Calendar APIs) using natural language prompts, rather than requiring specific syntax - The GitHub MCP Server was shown executing a command to create an issue, demonstrating this tool interaction capability. ### Conclusion - The key to responsible AI coding is maintaining control and engagement, rather than passively accepting AI output, ensuring developers review and understand the code and system interactions. ![Screenshot at 00:02: Announcement slide for GitHub Universe 25, scheduled for Oct 28-29, 2025, in San Francisco & Virtual.](https://ss.rapidrecap.app/screens/PvdptUZ3XeU/00-00-02.png) ![Screenshot at 00:15: Slide showing the title "How to Vibe Code Responsibly With a little help from MCPs" and speaker Rizel Scarlett, Tech Lead / Block.](https://ss.rapidrecap.app/screens/PvdptUZ3XeU/00-00-15.png) ![Screenshot at 00:49: Slide detailing the dangers of 'Vibe Coding': Insecure code, Poor Scalability, and Lose all your code, with accompanying illustrative icons.](https://ss.rapidrecap.app/screens/PvdptUZ3XeU/00-00-49.png) ![Screenshot at 01:59: Slide listing the 'Positives' of AI coding: Non-engineers can bring visions to life, and Engineers can build FASTER.](https://ss.rapidrecap.app/screens/PvdptUZ3XeU/00-01-59.png) ![Screenshot at 08:55: The screen transitions to an 'MCP UI Demo' slide, signaling the start of a practical demonstration.](https://ss.rapidrecap.app/screens/PvdptUZ3XeU/00-08-55.png)