Rubber Duck Thursdays - Building an MCP server

Key Concepts

• Model Context Protocol (MCP): A standard for AI tools (like GitHub Copilot) to access external services and tools.
• MCP Client: The AI tool using the MCP server (e.g., GitHub Copilot in VS Code).
• MCP Server: Provides access to tools and data for the AI tool.
• Tools: Actions the MCP server can perform (e.g., playing tic-tac-toe, creating a pull request).
• Resources: Data the MCP server can provide (e.g., the current state of a game board).
• Prompts: Example prompts shipped with the MCP server to help users effectively use the tools.
• gh.io/love-of-code: A summer-long hackathon for developers to build projects.

GitHub Change Log Review

• Consistent Tab Width Preferences: GitHub now consistently maintains user-defined tab width preferences for improved accessibility and code readability.
• Enhancements to Last Activity At in User Management API: Updates to the Copilot API to enhance the durability of the "last activity at" field.
• Secret Scanning Updates: Secret scanning now includes validity checks for Doppler, Mitrans, Onfido, Postman, and Segment.
• Dependabot Updates:
  • Expanded cooldown feature to support NuGet and Helm, allowing configuration of a minimum age requirement before Dependabot creates PRs for newly released dependencies.
  • Now supports newer versions of package managers across multiple ecosystems.
  • Batched updates are now generally available, allowing dismissal or reopening of multiple Dependabot alerts at once.
• GitHub Enterprise Importer: Updated IP addresses for GitHub Enterprise Importer.
• Copilot Coding Agent Updates:
  • Keeps pull request titles and bodies up to date as it addresses feedback.
  • Custom setup steps are more reliable and easier to debug, with progress now visible in the agent session logs.

Introduction to Model Context Protocol (MCP)

• MCP enables AI tools to interact with external services and tools, such as GitHub issues, Playwright for UI testing, and external APIs.
• It provides a consistent standard for plugging in these services, avoiding the need for separate standards for each AI tool.
• MCP consists of a client (the AI tool) and a server (providing access to tools and data).
• The user interacts with the AI agent (e.g., Copilot), which can then call MCP servers to access additional capabilities.

Turn-Based Game MCP Server: Show and Tell

• The presenter built a turn-based game (tic-tac-toe and rock-paper-scissors) as a demonstration of MCP.
• The project includes an MCP server, shared libraries, a web app, and web APIs.
• The web app allows users to play against an AI opponent, with Copilot orchestrating the AI's moves through the MCP server.
• The MCP server defines tools for playing games, analyzing game state, and waiting for player moves.
• The AI's moves are determined by coded algorithms within the MCP server, not by the large language model itself.

MCP Server Architecture and Concepts

• Tools: Actions the MCP server can perform. Examples include playing tic-tac-toe, creating a GitHub issue, or creating a pull request.
• Resources: Data the MCP server can provide. In the game app, this includes the current state of the game board or the last move made by a player.
• Prompts: Example prompts shipped with the MCP server to help users effectively use the tools.
• The MCP server is registered in Visual Studio Code using an mcp.json file, which specifies the location of the server's binaries.
• The server defines resources using URIs, which are translated into API calls to retrieve data.
• The presenter uses TypeScript and JavaScript for the project, leveraging existing SDKs for MCP.

MCP Server Implementation Details

• The server.ts file defines the capabilities of the MCP server, including resources, prompts, and tools.
• Tools are defined with descriptions, input schemas, and handlers that execute specific actions when the tool is invoked.
• The handleToolCall method maps tool names to specific handlers, which then execute the corresponding code.
• The readGameResource method retrieves the state of a game by calling a backend API.
• The AI's moves in tic-tac-toe are determined by algorithms that implement easy, medium, and hard difficulty levels.

Prompts and User Interaction

• Prompts can be shipped as part of the MCP server to provide users with guidance on how to use the tools effectively.
• Prompts are accessed in VS Code using the / command, followed by the MCP server name and the prompt name.
• The presenter demonstrates how to use prompts to get strategy advice for playing tic-tac-toe against different difficulty levels.
• Copilot can automatically discover and use the tools defined in the MCP server, or users can call the tools directly using specific commands.

Limitations and Considerations

• There is a limit to the number of tools that can be visible to Copilot (128).
• Too much context (too many tools, long prompts) can negatively impact the LLM's ability to understand and use the tools effectively.
• Authentication is important for MCP servers that access sensitive or private data.
• The choice of large language model can impact the effectiveness of tool calling.

Conclusion

The presenter demonstrated how MCP can be used to extend the capabilities of AI tools like GitHub Copilot. The turn-based game example provided a visual way to understand how MCP servers can provide access to tools, resources, and prompts. The key takeaway is that MCP enables developers to integrate their own code and services into AI workflows, allowing for more powerful and customized AI experiences. The presenter plans to open-source the project as a learning resource for the community.