Proposal: Integrate existing MCP Servers in Giselle

[satococoa]

What we want to achieve

We want to enable the use of MCP servers in Giselle. Specifically, we aim to enhance Giselle's functionality by integrating the official github-mcp-server and other useful MCP servers (Slack, filesystem, Brave Search, etc.).

Barriers

There are several key barriers to integrating existing MCP servers into Giselle:

1. Remote execution issues:

   • Most current MCP servers are designed for local execution ([Stdio](https://modelcontextprotocol.io/specification/2025-03-26/basic/transports/#stdio)) and are not suitable for remote execution
   • The [2025/03/26 specification](https://modelcontextprotocol.io/specification/2025-03-26/index) changes the communication method from SSE to [Streamable HTTP](https://modelcontextprotocol.io/specification/2025-03-26/basic/transports/#streamable-http), indicating that the specification itself is still actively evolving
   • Many SDKs, including the TypeScript SDK, may not have kept up with the latest specifications

2. Authentication and authorization challenges:

   • The MCP [authorization specification](https://modelcontextprotocol.io/specification/2025-03-26/basic/authorization/) has also been updated, and many servers have not yet implemented OAuth2-based authorization flows
   • Giselle already has GitHub authentication and holds Access Tokens, so we need to consider how to utilize these

3. Human in the loop:

   • MCP [tool execution](https://modelcontextprotocol.io/specification/2025-03-26/server/tools/) recommends human approval (SHOULD), and we need to consider how to implement this in Giselle
   • This is especially important to consider for execution during Preview and Integrate stages

Implementation approaches

I have considered two main implementation methods:

1. Adapter Development Approach

• Develop an adapter that converts Stdio servers to SSE servers, inspired by [supergateway](https://github.com/supercorp-ai/supergateway)
• Implement a [Custom Transport](https://modelcontextprotocol.io/specification/2025-03-26/basic/transports/#custom-transports) that can inject Giselle's authentication information (GitHub Tokens, etc.) for each session
• Challenges:
  • Existing MCP implementations use various languages (Python, TypeScript, Go, etc.) and would require infrastructure beyond Vercel
  • Hosting services like [Supermachine](https://supermachine.ai/) were considered but may not align well with Giselle's use case

2. Custom Implementation Approach

• Directly implement GitHub, Slack, and other necessary functions as Tool Calling in Giselle's AI SDK
• Prioritize user-needed functionality over MCP compatibility
• Achievable with reasonable development effort and easier to integrate with existing Giselle systems

Conclusion

While remote MCP server integration would be technically feasible, it presents a significant development challenge at this time. Additionally, the MCP specification itself is still actively evolving.

Therefore, directly implementing tools like GitHub and Slack integration into Giselle's AI SDK is likely the more practical approach. This allows us to deliver valuable functionality to users more quickly while leaving the door open to reconsider full MCP integration when the ecosystem matures.