Description

My team won the Best Use of Solana for TAMU Datathon 25: A Cold Case! This is how we did it.

Idea

We wanted to maximize our categories for prizes so we chose to double dip into the Solana and MCP server tracks. I had never used an MCP server nor deployed Solana programs so this was a great learning opportunity.

The core idea of this project was to increase accessibility to the Solana blockchain with an interface connecting natural language with the Solana API.

Essentially, the user could directly interact with the live blockchain just by describing their actions in natural language such as "create a wallet", "fund that wallet", "check that wallet", etc...

This tech stack included:

Claude Desktop as the AI Agent
Solana MCP as the MCP Server using web3.js and TypeScript
NodeJS as the runtime environment
Solana DevNet as the RPC API

This is the call flow:

Claude Desktop -> MCP Server (over stdio)
MCP Server -> Solana RPC (over HTTPS with JSON-RPC)
Solana RPC -> MCP Server (over HTTPS with JSON-RPC)
MCP Server -> Claude Desktop (over stdio)

Claude Desktop is responsible for translating natural language to function calls defined by the MCP server. The MCP server then uses those function calls to send predetermined RPC API requests to the live blockchain.

Future

In the future, this project can be expanded by including open source tooling. This would affect the AI agent and the communication protocol.

The new call flow would look something like this:

React Web App -> AI Agent (over WebSocket)
AI Agent -> ExpressJS Adapter (over HTTP REST)
ExpressJS Adapter -> MCP Server (over stdio)
MCP Server -> Solana RPC (over HTTPS with JSON-RPC)
Solana RPC -> MCP Server (over HTTPS with JSON-RPC)
MCP Server -> ExpressJS Adapter (over stdio)
ExpressJS Adapter -> AI Agent (over HTTP REST)
AI Agent -> React Web App (over WebSocket)