automcp

🚀 Overview

automcp allows you to easily convert tools, agents and orchestrators from existing agent frameworks into MCP servers, that can then be accessed by standardized interfaces via clients like Cursor and Claude Desktop.

We currently support deployment of agents, tools, and orchestrators as MCP servers for the following agent frameworks:

1. CrewAI
2. LangGraph
3. Llama Index
4. OpenAI Agents SDK
5. Pydantic AI
6. mcp-agent

🔧 Installation

Install from PyPI:

# Basic installation
pip install naptha-automcp

# UV
uv add naptha-automcp

Or install from source:

git clone https://github.com/napthaai/automcp.git
cd automcp
uv venv 
source .venv/bin/activate
pip install -e .

🧩 Quick Start

Create a new MCP server for your project:

Navigate to your project directory with your agent implementation:

cd your-project-directory

Generate the MCP server files via CLI with one of the following flags (crewai, langgraph, llamaindex, openai, pydantic, mcp_agent):

automcp init -f crewai

Edit the generated run_mcp.py file to configure your agent:

# Replace these imports with your actual agent classes
from your_module import YourCrewClass

# Define the input schema
class InputSchema(BaseModel):
    parameter1: str
    parameter2: str

# Set your agent details
name = "<YOUR_AGENT_NAME>"
description = "<YOUR_AGENT_DESCRIPTION>"

# For CrewAI projects
mcp_crewai = create_crewai_adapter(
    orchestrator_instance=YourCrewClass().crew(),
    name=name,
    description=description,
    input_schema=InputSchema,
)

Install dependencies and run your MCP server:

automcp serve -t sse

📁 Generated Files

When you run automcp init -f <FRAMEWORK>, the following file is generated:

run_mcp.py

This is the main file that sets up and runs your MCP server. It contains:

• Server initialization code
• STDIO and SSE transport handlers
• A placeholder for your agent implementation
• Utilities to suppress warnings that might corrupt the STDIO protocol

You'll need to edit this file to:

• Import your agent/crew classes
• Define your input schema (the parameters your agent accepts)
• Configure the adapter with your agent

🔍 Examples

Running the examples

The repository includes examples for each supported framework:

# Clone the repository
git clone https://github.com/NapthaAI/automcp.git
cd automcp

# Install automcp in development mode
pip install -e .

# Navigate to an example directory
cd examples/crewai/marketing_agents

# Generate the MCP server files (use the appropriate framework)
automcp init -f crewai

# Edit the generated run_mcp.py file to import and configure the example agent
# (See the specific example's README for details)

# Add a .env file with necessary environmental variables

# Install dependencies and run
automcp serve -t sse

Each example follows the same workflow as a regular project:

1. Run automcp init -f <FRAMEWORK> to generate the server files
2. Edit run_mcp.py to import and configure the example agent
3. Add a .env file with necessary environmental variables
4. Install dependencies and serve using automcp serve -t sse

CrewAI example

Here's what a typical configured run_mcp.py looks like for a CrewAI example:

import warnings
from typing import Any
from automcp.adapters.crewai import create_crewai_adapter
from pydantic import BaseModel
from mcp.server.fastmcp import FastMCP

mcp = FastMCP("MCP Server")

warnings.filterwarnings("ignore")

from crew import MarketingPostsCrew

class InputSchema(BaseModel):
    project_description: str
    customer_domain: str

name = "marketing_posts_crew"
description = "A crew that posts marketing posts to a social media platform"

# Create an adapter for crewai
mcp_crewai = create_crewai_adapter(
    orchestrator_instance=MarketingPostsCrew().crew(),
    name=name,
    description=description,
    input_schema=InputSchema,
)
mcp.add_tool(
    mcp_crewai,
    name=name,
    description=description
)

# Server entrypoints
def serve_sse():
    mcp.run(transport="sse")

def serve_stdio():
    # Redirect stderr to suppress warnings that bypass the filters
    import os
    import sys

    class NullWriter:
        def write(self, *args, **kwargs):
            pass
        def flush(self, *args, **kwargs):
            pass

    # Save the original stderr
    original_stderr = sys.stderr

    # Replace stderr with our null writer to prevent warnings from corrupting STDIO
    sys.stderr = NullWriter()

    # Set environment variable to ignore Python warnings
    os.environ["PYTHONWARNINGS"] = "ignore"

    try:
        mcp.run(transport="stdio")
    finally:
        # Restore stderr for normal operation
        sys.stderr = original_stderr

if __name__ == "__main__":
    import sys
    if len(sys.argv) > 1 and sys.argv[1] == "sse":
        serve_sse()
    else:
        serve_stdio()

🔄 Running Your MCP Server

After setting up your files, you can run your server using one of these methods:

# Using the automcp CLI
automcp serve -t stdio    # STDIO transport
automcp serve -t sse      # SSE transport

# Or run the Python file directly
python run_mcp.py       # STDIO transport
python run_mcp.py sse   # SSE transport

# Or with uv run (if configured in pyproject.toml)
uv run serve_stdio
uv run serve_sse

Note about transport modes:

• STDIO: You don't need to run the server manually - it will be started by the client (Cursor)
• SSE: This is a two-step process:
  1. Start the server separately: python run_mcp.py sse or automcp serve -t sse
  2. Add the mcp.json configuration to connect to the running server

If you want to use the uv run commands, add the following to your pyproject.toml:

[tool.uv.scripts]
serve_stdio = "python run_mcp.py"
serve_sse = "python run_mcp.py sse"

☁️ Deploying with Naptha's MCPaaS

Naptha supports deploying your newly-created MCP server to our MCP servers-as-a-service platform! It's easy to get started.

Setup

Naptha's MCPaaS platform requires your repository be set up with uv. This means you need a couple configurations in your pyproject.toml.

First, make sure the run_mcp.py file generated by naptha-automcp is the root of your repository.

Second, make sure your pyproject.toml has the following configurations:

[build-system]
requires = [ "hatchling",]
build-backend = "hatchling.build"

[project.scripts]
serve_stdio = "run_mcp:serve_stdio"
serve_sse = "run_mcp:serve_sse"

[tool.hatch.metadata]
allow-direct-references = true

[tool.hatch.build.targets.wheel]
include = [ "run_mcp.py",]
exclude = [ "__pycache__", "*.pyc",]
sources = [ ".",]
packages = ["."]

If your agent is in a subdirectory / package of your repository:

pyproject.toml
run_mcp.py
my_agent/
|---| __init__.py
    | agent.py

Make sure that it's imported like this in run_mcp.py:

from my_agent.agent

Not like below, since this will cause the build to fail:

from .my_agent.agent

Once you have configured everything, commit and push your code (but not your environment variables!) to github. Then, you can test it to make sure you set up everything correctly:

uvx --from https://github.com/your-username/your-repo serve_sse

If this results in your MCP server being launched on port 8000 successfully, you're good to go!

Launching your server

1. go to labs.naptha.ai
2. Sign in with your github account
3. Pick the repository you edited from your repository list -- we autodiscover your github repos.
4. add your environment variables e.g. OPENAI_API_KEY, etc.
5. Click Launch.
6. Copy the SSE URL, and paste it into your MCP client:

🔌 Using with MCP Clients

Cursor

To integrate with Cursor IDE, create a .cursor folder in your project root and add an mcp.json file with the following configuration:

{
    "mcpServers": {
        "crew-name-stdio": {
            "type": "stdio",
            "command": "/absolute/path/to/your/.venv/bin/uv",
            "args": [
                "--directory",
                "/absolute/path/to/your/project_dir",
                "run",
                "serve_stdio"
            ],
            "env": {
                "OPENAI_API_KEY": "sk-",
                "SERPER_API_KEY": ""
            }
        },
        
        "crew-name-python": {
            "type": "stdio",
            "command": "/absolute/path/to/your/.venv/bin/python",
            "args": [
                "/absolute/path/to/your/project_dir/run_mcp.py"
            ],
            "env": {
                "OPENAI_API_KEY": "sk-",
                "SERPER_API_KEY": ""
            }
        },
        
        "crew-name-automcp": {
            "type": "stdio",
            "command": "/absolute/path/to/your/.venv/bin/automcp",
            "args": [
                "serve",
                "-t",
                "stdio"
            ],
            "cwd": "/absolute/path/to/your/project_dir",
            "env": {
                "OPENAI_API_KEY": "sk-",
                "SERPER_API_KEY": ""
            }
        },
        
        "crew-name-sse": {
            "type": "sse",
            "url": "http://localhost:8000/sse"
        }
    }
}

Note: Be sure to replace all placeholder paths with absolute paths to your actual files and directories.

Direct GitHub Execution

Push your project to GitHub and use:

{
   "mcpServers": {
       "My Agent": {
           "command": "uvx",
           "args": [
               "--from",
               "git+https://github.com/your-username/your-repo",
               "serve_stdio"
           ],
           "env": {
               "OPENAI_API_KEY": "your-key-here"
           }
       }
   }
}

🛠️ Creating New Adapters

Want to add support for a new agent framework? Here's how:

1. Create a new adapter file in automcp/adapters/ (or add to an existing framework file):

# automcp/adapters/framework.py
import json
import contextlib
import io
from typing import Any, Callable, Type
from pydantic import BaseModel

def create_framework_adapter(
    agent_instance: Any,
    name: str,
    description: str,
    input_schema: Type[BaseModel],
) -> Callable:
    """Doc string for your function"""
    
    # Get the field names and types from the input schema
    schema_fields = input_schema.model_fields

    # Create the parameter string for the function signature
    params_str = ", ".join(
        f"{field_name}: {field_info.annotation.__name__}"
        for field_name, field_info in schema_fields.items()
    )

    # Create the function body that constructs the input schema
    # Note: You may need to adjust the method calls (kickoff, model_dump_json)
    # to match your framework's specific API
    body_str = f"""def run_agent({params_str}):
        inputs = input_schema({', '.join(f'{name}={name}' for name in schema_fields)})
        with contextlib.redirect_stdout(io.StringIO()):
            result = agent_instance.framework_specific_run(inputs=inputs.model_dump())
        return result.framework_specific_result()
    """

    # Create a namespace for the function
    namespace = {
        "input_schema": input_schema,
        "agent_instance": agent_instance,
        "json": json,
        "contextlib": contextlib,
        "io": io,
    }

    # Execute the function definition in the namespace
    exec(body_str, namespace)

    # Get the created function
    run_agent = namespace["run_agent"]

    # Add proper function metadata
    run_agent.__name__ = name
    run_agent.__doc__ = description

    return run_agent

2. Create an example in examples/your_framework/

📝 Notes

• When working with STDIO transport, be careful with print statements in your agent code as they can corrupt the protocol
• The MCP Inspector can be used for debugging: npx @modelcontextprotocol/inspector
• Remember that for STDIO mode, the client (like Cursor) will start the server for you
• For SSE mode, you need to manually start the server and then configure the client to connect to it