Build your MCP Server with One Prompt

Roo Code makes MCP integration easier than ever, allowing you to create your own MCP servers with LLMs directly in your IDE. Roo Code is an editor-based AI Assistant, with various ‘modes’ enabling you to personalize your coding experience. Roo fits into your workflow, and allows you to easily build tools to enhance it.

Importance of MCP Servers

We’ve established in previous blogs how important MCP servers are in the development process. They address core limitations for LLMs, such as lack of relevant context, and an inability to interact with your workflow. Having the right MCP servers for your AI Assistant can greatly improve LLM productivity, transforming your assistant from a simple code generator to a powerful architect. For a list of especially useful MCP servers, check out our blog.

Limitations of MCP

The issue many developers run into, though, is that they have very specific use cases that can’t be addressed by existing MCP servers. Imagine you’re trying to develop a Slackbot that messages a channel every time your deployment finishes and tests pass. Something like that is not something you can leverage a premade MCP server for; it’s too nuanced and specialized.

Roo Code’s offering

When you need these custom MCP servers to query a private API or to trigger something like a Discord notification, Roo Code can be a big help. They have a feature where you can use Roo to create a custom MCP server addressing whatever you need. As an example, you could type something like: "Create an MCP tool that checks the current weather using the OpenWeather API." Roo will handle the scaffolding, code generation, and registration for you. You can find full instructions on how to set this up in their documentation.

Building Useful MCP Servers

While Roo Code makes building custom MCP servers a whole lot easier, building helpful ones remains difficult. Below are some tips on how to create MCP servers that are LLM-friendly, and that you’ll actually use.

Keep tools small in scope

Instead of having one MCP server handle all of your requirements, each tool should have one clear job. This prevents confusion on both the LLM and human side, making your servers easier to maintain and understand.

Handle Errors Eloquently

Debugging MCP servers can be a pain, but you can make it much easier by properly wrapping calls in try/catch blocks, returning structured error messages that show up directly in chat. This is especially powerful because Roo getting in a feedback loop greatly improves how quickly you can iterate and how reliable your tools are.

Test Tools Before Connecting

Before hoping that Roo will immediately tap into your new tool, check that it works correctly with curl, or node server.js. If it doesn’t work manually, it won’t work through MCP.

Name Tools Well

Although it seems somewhat obvious, naming your tools well (outlined in this article) helps both you and your model remember what each server does. Roo relies on tool names to decide which one to call, so names like get_exact_time or post_to_discord are much more useful than tool_7.

Final Thoughts

MCP Servers are a great tool to transform your AI Assistant from a passive recommendation engine to an active helper. With custom MCP servers, your AI teammate can fetch metrics, query internal APIs, and automate anything you can set your mind to! If you’ve been wondering how you can integrate personal tools into your development workflow, this is it!

When you build something cool, share it! We have a Discord server where we’d love to hear your discoveries.