PyMCP
A simple, tiny, and asynchronous server and client implementation for the Modern Context Protocol that allows you to easily expose functions as remote tools using a decorator with hot-reloading capability.
README Documentation
PyMCP
PyMCP is a simple, tiny, and asynchronous server and client implementation for the Modern Context Protocol (MCP).
Features
- Simple Tool Definition: Easily expose functions as remote tools using a simple
@tool
decorator. - Hot-Reloading: Tools can be added, removed, or modified on the fly, and the server will hot-reload them without a restart.
Quick Start
Running the Server
-
Installation (from source):
git clone <repository-url> cd pymcp pip install -r requirements.txt pip install -e .
-
Run the Server:
- run with default settings:
pymcp
- run with custom settings:
pymcp --host 0.0.0.0 --port 9000 --tool-repo ./my_tools --log-level DEBUG
Using the Client
import asyncio
import pymcp
async def main():
try:
async with pymcp.Client("localhost", 8765) as client:
# Ping the server
pong = await client.call("ping")
# Call the custom 'add' tool
result = await client.call("add", a=5, b=7)
print(f"5 + 7 = {result}")
# Discover available tools
tools = await client.call("list_tools_available")
print("Available tools:", tools)
except Exception as e:
print(f"An error occurred: {e}")
if __name__ == "__main__":
asyncio.run(main())
Configuration
- Arguments passed to start_server().
- Command-line arguments (e.g., --port).
- Environment variables (e.g., PYMCP_PORT=9000).
- Values in a .env file in your project's root.
- Default values in the Settings class.
all configuration options:
# .env
# Server network settings
# PYMCP_HOST=<your_host_address>
PYMCP_HOST=127.0.0.1
# PYMCP_PORT=<your_port_number>
PYMCP_PORT=8765
# List of paths to user-defined tool directories.
# Paths should be separated by commas.
# absolute or relative paths
# Example: /path/to/my_tools,/another/path/for_tools
PYMCP_USER_TOOL_REPOS=./my_custom_tools,../shared_tools,/home/user/tools
# Logging level
# Valid values: DEBUG, INFO, WARNING, ERROR, CRITICAL
PYMCP_LOG_LEVEL=INFO
Server
request->response flow
client -> connection manager -> validator -> router -> tool executor -> return result -> client
create a tool
import pymcp
@pymcp.tool
def add(a: int, b: int) -> int:
"""
Add two integer numbers.
argument a: First integer number.
argument b: Second integer number.
return: The sum of the two numbers.
"""
return a + b
@pymcp.tool
def retrieve_data() -> str:
"""
Retrieve user data from source
no arguments is needed
"""
return data
Tool Repository discovery
The server discovers tools by scanning all .py
files within the directories specified by the --tool-repo
CLI argument or the PYMCP_TOOL_REPOS
environment variable.
You can organize your tools into multiple files and directories. The loader will scan them recursively.
example directory structure:
my_tools/
├── __init__.py
├── math_tools.py
├── string_tools.py
└── data_tools/
├── __init__.py
├── user_data.py
└── system_data.py
second_tools/
├── __init__.py
├── network_tools.py
└── file_tools.py
Hot-Reloading
The server watches the tool repositories for file changes. If you add, modify, or delete a Python file in a tool directory, the server will automatically perform a reload:
- It rebuilds the entire tool registry from scratch.
- It atomically swaps the old registry with the new one.
This allows you to update tool logic on a live server without a restart.
PyMCP supports a simple form of dependency injection. If a tool function's signature includes a parameter named tool_registry
, the server will automatically provide the ToolRegistry
instance to it at execution time.
Client
Connect to the server
import pymcp
async def main():
async with pymcp.Client("localhost", 8765) as client:
result = await client.call("ping")
print(result)
request execute tools
import pymcp
async def main():
async with pymcp.Client("localhost", 8765) as client:
result = await client.call("add", a=5, b=7)
print(f"5 + 7 = {result}")
tools = await client.call("list_tools_available")
print("Available tools:", tools)
Protocol
PyMCP uses a simple, JSON-based messaging protocol over a standard WebSocket connection.
- Client-to-Server: Requests
{
"header": {
"correlation_id": "a1b2c3d4-e5f6-7890-1234-567890abcdef"
},
"body": {
"tool": "add",
"args": {
"a": 5,
"b": 10
}
}
}
- Server-to-Client: Responses (Success)
{
"status": "success",
"header": {
"correlation_id": "a1b2c3d4-e5f6-7890-1234-567890abcdef"
},
"body": {
"tool": "add",
"result": 15
},
"error": null
}
- Server-to-Client: Responses (Error)
{
"status": "error",
"header": {
"correlation_id": "a1b2c3d4-e5f6-7890-1234-567890abcdef"
},
"body": null,
"error": {
"code": "execution_error",
"message": "An unexpected error occurred while executing tool 'add'."
}
}
Design
PyMCP is designed to be simple, extensible, and easy to use. The server itself is just a WebSocket server that handles incoming requests, validates them, routes them to the appropriate tool executor, and returns the results.
No concept like "resources", "tools", or "prompts", since all of them are just a function that input something and return something. Everything is a tool. It leaves to user defining the scope of the tools.
The internal tools like list_tools_available
and ping
, are also some type of "tools". Tools and Server are decoupled. core tools are like extension of the server.