An MCP-Server that reads the contents of Python packages. This Server allows Large Language Models (LLMs) and other AI Agents to inspect the contents of Python packages in a specified environment.
pypreader-mcp
acts as a bridge between AI models and the local Python environment. By exposing a set of tools via MCP, it enables AI to programmatically browse installed packages, view their file structures, and read their source code. This is useful for tasks such as code analysis, dependency checking, and automated programming assistance.
When I use AI-integrated programming IDEs like Cursor or Trae, I always find that the currently used model is unaware of the third-party libraries I need.
But sometimes they pretend to know and seriously generate a bunch of indescribable stuff; other times they search the internet, but the results are mostly not good, making it hard to find any useful information.
So I created this MCP service. It can read documentation from the official website pypi.org or read source code from the site-packages environment corresponding to your Python, allowing for a more direct understanding of the contents of the third-party libraries you want to use.
This Server provides the following tools to MCP clients:
get_pypi_description(package_name: str)
: Retrieve the official description of a package from PyPI.get_package_directory(package_name: str)
: List the entire file and directory structure of a specified installed package.get_source_code_by_path(package_path: str)
: Retrieve the complete source code of a specific file within a package.get_source_code_by_symbol(package_path: str, symbol_name: str)
: Get the definition (code snippet) of a specified symbol (function, class, etc.).
This tool is designed to act as an MCP Server for use in AI-based environments such as Cursor or Trae.
In the MCP Server configuration of your AI environment, add a new Server with the following settings. This allows the AI to run the service directly from its Git repository using uvx
.
{
"mcpServers": {
"PypReader-MCP": {
"command": "uvx",
"args": [
"--from",
"git+https://github.com/zakahan/pypreader-mcp.git",
"pypreader-mcp"
],
"env": {
"CURRENT_PYTHON_PATH": "<your-python-path>",
"CURRENT_LOGGING_LEVEL": "INFO"
}
}
}
}
When configuring the MCP Server in your AI environment, you can specify the following env parameters:
CURRENT_PYTHON_PATH
: Specifies the path to the Python executable of the target package installation environment. If not provided, it defaults to the Python executable running the Server. You can find the correct path by activating your project's Python environment and runningwhich python
in the terminal.CURRENT_LOGGING_LEVEL
: Sets the logging level of the Server. Options areDEBUG
,INFO
,WARNING
,ERROR
,CRITICAL
. The default value isINFO
.
If you use a Python virtual environment to configure your Python project, you typically need to modify python_path
as needed to switch to your specified Python environment.
Take Trae as an example. As of now (2025-07-02), the doubao-seed-1.6 model is unaware of the fastmcp
package (in fact, most models don't recognize it either). Under normal circumstances, it would either pretend to know and output a bunch of messy, indescribable incorrect content (even thinking I'm talking about FastAPI), or perform a clumsy search and find all sorts of disorganized information.
This time, I created a Trae agent equipped with this project's mcp-server. The result is as follows: Trae can understand my project and then write a fastmcp
service to complete my task.
This is the entire process of Trae completing my requested task
These are the specific details of the tool calls
This project is licensed under the MIT License. For details, please refer to the LICENSE file.
- 2025-07-06: Rewrote
get_source_code_by_symbol
to use reading and writing temporary files instead of stdio to support the Windows platform. - 2025-07-04: Rewrote the
get_source_code_by_symbol
tool to fix the issue where it couldn't read classes or functions belonging to sub-packages.
- Resolve the issue where
python-package-name
is inconsistent with the actual path, e.g., the package namegoogle-adk
has an actual path ofgoogle/adk
and is actually imported asgoogle.adk
- Design a suitable prompt for the Agent. (The Trae version is here)