MCP Tools

The Model Context Protocol (MCP) is an open standard for connecting AI models to external tools and data sources. Elsai has first-class MCP support — connect any MCP server in seconds.

What is MCP?

MCP servers expose tools, resources, and prompts over a standardised protocol. Thousands of pre-built MCP servers exist for databases, APIs, file systems, browsers, and more.

Basic usage

from elsai import Agent
from elsai.tools.mcp import MCPClient
from mcp import stdio_client, StdioServerParameters

# Connect to an MCP server via stdio
client = MCPClient(
    lambda: stdio_client(StdioServerParameters(
        command="uvx",
        args=["mcp-server-fetch"]
    ))
)

with client:
    tools = client.list_tools_sync()
    agent = Agent(tools=tools)
    result = agent("Fetch the content of https://example.com")
    print(result)

Async context manager

import asyncio
from elsai import Agent
from elsai.tools.mcp import MCPClient
from mcp import stdio_client, StdioServerParameters

async def main():
    client = MCPClient(
        lambda: stdio_client(StdioServerParameters(
            command="uvx",
            args=["mcp-server-filesystem", "/tmp"]
        ))
    )

    async with client:
        tools = await client.list_tools()
        agent = Agent(tools=tools)
        result = await agent.invoke_async("List all files in the /tmp directory")
        print(result)

asyncio.run(main())

HTTP/SSE transport

Connect to an MCP server running as an HTTP service:

from elsai.tools.mcp import MCPClient
from mcp.client.sse import sse_client

client = MCPClient(
    lambda: sse_client("http://localhost:8080/sse")
)

with client:
    agent = Agent(tools=client.list_tools_sync())
    agent("Use the remote server tools")

AWS Documentation MCP server

from elsai import Agent
from elsai.tools.mcp import MCPClient
from mcp import stdio_client, StdioServerParameters

aws_docs = MCPClient(
    lambda: stdio_client(StdioServerParameters(
        command="uvx",
        args=["awslabs.aws-documentation-mcp-server@latest"]
    ))
)

with aws_docs:
    agent = Agent(tools=aws_docs.list_tools_sync())
    result = agent("How do I use Amazon Bedrock with Python?")

Combining MCP tools with function tools

from elsai import Agent, tool
from elsai.tools.mcp import MCPClient
from mcp import stdio_client, StdioServerParameters

@tool
def my_custom_tool(x: str) -> str:
    """A custom Python tool.

    Args:
        x: Input value.
    """
    return x.upper()

client = MCPClient(
    lambda: stdio_client(StdioServerParameters(command="uvx", args=["mcp-server-fetch"]))
)

with client:
    agent = Agent(tools=[my_custom_tool, *client.list_tools_sync()])
    agent("Use both custom and MCP tools")

Filtering MCP tools

Only expose a subset of an MCP server's tools:

all_tools = client.list_tools_sync()
filtered = [t for t in all_tools if t.name in {"read_file", "write_file"}]
agent = Agent(tools=filtered)

Multiple MCP servers

from elsai.tools.mcp import MCPClient
from mcp import stdio_client, StdioServerParameters

db_client = MCPClient(lambda: stdio_client(StdioServerParameters(command="uvx", args=["mcp-server-postgres", "--url", DATABASE_URL])))
fs_client = MCPClient(lambda: stdio_client(StdioServerParameters(command="uvx", args=["mcp-server-filesystem", "."])))

with db_client, fs_client:
    tools = [*db_client.list_tools_sync(), *fs_client.list_tools_sync()]
    agent = Agent(tools=tools)
    agent("Read the schema from the database and write it to a file")

MCP resources

Access MCP resources (not just tools):

async with client:
    resources = await client.list_resources()
    for resource in resources:
        print(f"{resource.name}: {resource.uri}")

    content = await client.read_resource("file:///path/to/file.txt")

Popular MCP servers

Package	What it does
mcp-server-fetch	Fetch web pages
mcp-server-filesystem	Read/write local files
mcp-server-postgres	Query PostgreSQL
mcp-server-sqlite	Query SQLite
mcp-server-github	GitHub API
awslabs.aws-documentation-mcp-server	AWS docs
mcp-server-brave-search	Brave web search

Install any MCP server with uvx <package> (no installation needed) or pip install <package>.