Model Context Protocol (MCP)

Alith supports Model Context Protocol (MCP), a technique that enables AI models to dynamically integrate external services and data sources during inference. This protocol acts as a bridge between LLMs and various tools/APIs, allowing models to:

Access real-time context from external systems (GitHub, databases, APIs, etc.).
Extend capabilities without retraining by connecting to specialized tools.
Maintain fresh knowledge through live data connections.
Execute complex workflows via chained service integrations.

The protocol operates through lightweight server adapters defined in the configuration file. Each MCP server:

Runs as a separate process using native runtime (Node.js/Python).
Implements standardized interfaces for tool discovery and execution.
Can chain multiple services through semantic routing.
Maintains security through process isolation.

Rust

Read MCP Servers from Config


use alith::{Agent, Chat, LLM};
 
#[tokio::main]
async fn main() -> Result<(), anyhow::Error> {
    let model = LLM::from_model_name("gpt-4")?;
    let agent = Agent::new("simple agent", model)
        .preamble("You are a calculator here to help the user perform arithmetic operations. Use the tools provided to answer the user's question.")
        .mcp_config_path("servers_config.json").await?;
    let response = agent.prompt("Calculate 10 - 3").await?;
    println!("{}", response);
    Ok(())
}

An example of MCP config servers is as follows


{
    "mcpServers": {
        "github": {
            "command": "npx",
            "args": [
                "-y",
                "@modelcontextprotocol/server-github"
            ]
        }
    }
}

Note: Note that most MCP servers are written in Python or Node.js, you need to install the Python package manage tool uv and the Node.js package manage tool npx

Python

Read MCP Servers from Config


from alith import Agent
 
agent = Agent(
    name="Calculator Agent",
    model="gpt-4o-mini",
    preamble="You are a calculator here to help the user perform arithmetic operations. Use the tools provided to answer the user's question.",
    mcp_config_path="servers_config.json",
)
print(agent.prompt("Calculate 10 - 3"))

An example of MCP config servers is as follows


{
    "mcpServers": {
        "github": {
            "command": "npx",
            "args": [
                "-y",
                "@modelcontextprotocol/server-github"
            ]
        }
    }
}

Note: Note that most MCP servers are written in Python or Node.js, you need to install the Python package manage tool uv and the Node.js package manage tool npx

Node.js

Read MCP Servers from Config


import { Agent } from 'alith'
 
const agent = new Agent({
  model: 'gpt-4',
  preamble:
    'You are a calculator here to help the user perform arithmetic operations. Use the tools provided to answer the user question.',
  mcpConfigPath: 'servers_config.json',
})
console.log(await agent.prompt('Calculate 10 - 3'))

An example of MCP config servers is as follows


{
    "mcpServers": {
        "github": {
            "command": "npx",
            "args": [
                "-y",
                "@modelcontextprotocol/server-github"
            ]
        }
    }
}

Note: Note that most MCP servers are written in Python or Node.js, you need to install the Python package manage tool uv and the Node.js package manage tool npx

Write an MCP Server

Of course, writing an MCP Server is also easy. Here, we take using Python to write a simple Web3 interactive MCP Server as an example.

Install the mcp dependency.


python3 -m pip install mcp -U

Write the MCP Server code named mcp_server.py.


from mcp.server.fastmcp import FastMCP
from web3 import Web3
 
# Initialize the MCP server
mcp = FastMCP("MetisBlockServer")
 
# Connect to Metis mainnet
w3 = Web3(Web3.HTTPProvider("https://andromeda.metis.io/?owner=1088"))
 
 
# MCP Tool: Fetch Latest Block Number (already working)
@mcp.tool()
def get_latest_block() -> int:
    """Fetches the latest block number from Metis mainnet."""
    return w3.eth.block_number
 
 
# MCP Tool: Fetch Previous Block Number (already working)
@mcp.tool()
def get_previous_block() -> int:
    """Fetches the previous block number from Metis mainnet."""
    latest_block = get_latest_block()
    return latest_block - 1
 
 
# Run the MCP server
if __name__ == "__main__":
    mcp.run()

Then you can append the following config to your MCP config file servers_config.json.


{
    "mcpServers": {
        "metis": {
            "command": "python3",
            "args": [
                "mcp_server.py"
            ]
        }
    }
}

At last, you can run the Alith agent code.


from alith import Agent
 
agent = Agent(
    model="gpt-4",
    mcp_config_path="servers_config.json",
)
print(agent.prompt("What is the latest block number on Metis mainnet?"))