My journey: Transitioning from connery-sdk to MCP

[machulav]

Pre-submission Checklist

• I have verified that this discussion would not be more appropriate as an issue in a specific repository
• I have searched existing discussions to avoid duplicates

Discussion Topic

Hi MCP Team,

First of all, I want to say that I really admire what you’re doing with the Model Context Protocol. I’ve been exploring the spec, and its potential is truly exciting!

Funny enough, I’ve been working on something similar over the past year for my company. It’s called connery-sdk, an open-source SDK designed to help build plugin servers (similar to MCP servers) with actions that act like MCP tools but can also serve as prompts and resources. The SDK is easy to integrate with various LLMs and tools, such as OpenAI, LangChain, and agents. LangChain even featured it in one of their blog posts.

Each plugin server is an API generated by the SDK, deployed to an AWS Lambda function, and accessed via API by clients (similar to MCP clients). We use it in our product to easily extend it with plugins. Although connery-sdk is much simpler than MCP and focused solely on tools, I noticed many similar design decisions between the two.

Before starting this project, I worked extensively with no-code platforms like Make and Zapier. I was inspired by their approach of creating small apps that integrate with third-party systems and act as micro backends to perform specific, useful tasks. However, all of them are closed-source, and there isn’t a strong community around them. On the other hand, I’ve also worked a lot with GitHub Actions, which are similar in terms of reusable actions but stand out because of their incredible community-driven model. I realized that AI would benefit from something similar—a collection of small, reusable actions in an open and community-driven format to foster development and adoption. That vision led me to create connery-sdk.

Unfortunately, I wasn’t able to gain much traction with it. Now, seeing how well MCP is coming together, I’ve decided to sunset connery-sdk in favor of MCP. And I’d love to explore how I could contribute to MCP’s growth.

We also plan to migrate our product from using connery-sdk to MCP once remote MCP connections become available.

That’s essentially what I wanted to share with you 😊
Thank you for the incredible work you’re doing. I’m excited to see where MCP goes!

Volo

[Zhaobudaoyuema]

Hey, I'm curious about how Connery-SDK handles the issue of adding many tools, which can make it hard for large models to call them precisely. I noticed that Claude seems to add a lot of MCP tools without this problem. How do they manage to do that?