Implement MCP Server as CLI that will be used to interact with Dust agents

[tdraier]

Context

An MCP server is a standardized protocol implementation that:

1. Exposes "tools" as endpoints that AI applications can call, with each tool defined by:

   • A unique name
   • Description
   • Input parameters (validated with schemas)
   • Output format

2. Uses stdio or SSE for communication, where:

   • stdio: reads from standard input, writes to standard output
   • SSE: server-sent events for streaming responses

3. Implements authentication flow, typically:

   • API key
   • OAuth for user authentication

4. Handles requests/responses in a standardized JSON-RPC format, allowing AI models to:

   • Discover available tools
   • Call specific functions
   • Receive structured responses

Implementation

In this task, we will implement a local MCP server, running as as CLI, that will be used to interact with Dust agents. As an example, the tools we'll provide can include :

• list_agents to expose available Dust agents
• ask_agent as main entry point for agent interactions

The client can be a standard MCP test client, or an IDE like Cursor that supports MCP.

We'll use stdio for communication, as we're running a local client and server.

MCP provides frameworks in different languages to implement the server. You can choose the language you feel more comfortable with, although typescript will allow you to use our own SDK to call the Dust API.

Resources

https://spec.modelcontextprotocol.io/specification/2024-11-05/
https://modelcontextprotocol.io/quickstart/server
https://www.apideck.com/blog/unlocking-ai-potential-how-to-quickly-set-up-a-cursor-mcp-server