CrewAI - OpenAI - Invalid schema for function

[RobertoSannino]

Description:

Trying to use an hosted docker neo4j-mcp server through the python CrewAI library with OpenAI model results in error when using the neo4j tools.

{'error': {'message': "Invalid schema for function 'read_cypher': None is not of type 'object', 'boolean'.", 'type': 'invalid_request_error', 'param': 'tools[1].function.parameters', 'code': 'invalid_function_parameters'}}

The library correctly reads the available mcp tools.

How to Reproduce:

Docke compose

Request:

services:
  neo4j-mcp:
    image: mcp/neo4j:latest
    container_name: neo4j-mcp
    ports:
      - "8000:8000"
    environment:
      - NEO4J_URI=neo4j+s://demo.neo4jlabs.com:7687
      - NEO4J_DATABASE=companies
      - READ_ONLY="true"
      - MAX_RESULT_ROWS=1000
      - NEO4J_TRANSPORT_MODE=http
      - NEO4J_TELEMETRY=false
      - NEO4J_MCP_HTTP_HOST=0.0.0.0
      - NEO4J_MCP_HTTP_PORT=8000
      - NEO4J_MCP_HTTP_PATH=/mcp/
      - NEO4J_MCP_HTTP_ALLOWED_ORIGINS="*"
      - NEO4J_MCP_HTTP_ALLOWED_HOSTS="*"
      - NEO4J_READ_TIMEOUT="300"

Python Code:

from crewai import LLM, LLM, Agent, Task, Crew
from crewai_tools import MCPServerAdapter
from crewai.mcp import MCPServerHTTP
from dotenv import load_dotenv

# Load .env file
load_dotenv('./env.env')

token = "Y29tcGFuaWVzOmNvbXBhbmllcw=="
server_parameters = {
    "url": "http://localhost:8000/mcp",
    "transport": "streamable-http",
    "headers": {
        "Authorization": f"Bearer {token}"
    }
}

# Optionally logging callbacks from Agents & Tasks
def log_step_callback(output):
    print(
        f"""
        Step completed!
        details: {output.__dict__}
    """
    )


def log_task_callback(output):
    print(
        f"""
        Task completed!
        details: {output.__dict__}
    """
    )


# Convenience for fastAPI call
def run(prompt: str):
    
    # Load the MCP Tools
    with MCPServerAdapter(server_parameters) as tools:

        print(f"Available tools from MCP server(s): {[tool.name for tool in tools]}")

        # Create an agent with access to tools
        mcp_agent = Agent(
            role="MCP Tool User",
            goal="Utilize tools from MCP servers.",
            backstory="I can connect to MCP servers and use their tools.",
            tools=tools,
            max_iterations=3,
            step_callback=log_step_callback,  # Optional
            llm = LLM(
                model="openai/gpt-4o",
                api_key="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",  # Or set OPENAI_API_KEY
                temperature=0.7,
                max_tokens=4000,
            )
        )

        # Create a task referrencing user prompt
        processing_task = Task(
            description="""Process the following prompt about the Neo4j graph database: {prompt}""",
            expected_output="A brief report on the outcome of the command: {prompt}",
            agent=mcp_agent,
            callback=log_task_callback,  # Optional
        )

        # Create the crew
        crew = Crew(agents=[mcp_agent], tasks=[processing_task], verbose=False)

        # Run the crew w/ the user prompt
        result = crew.kickoff(inputs={"prompt": prompt})

        # Return the final answer
        return {"result": result}


# For running as a script
if __name__ == "__main__":

    cmd = "What are the competitors of BigFix?"
    result = run(cmd)

    print(
        f"""
        Query completed!
        result: {result}
    """
    )

Expected Behavior:

For an MCP Server, the desiderata is that LLMs and libraries work smoothly with the defined tools.

Actual Behavior:

It seems that the tool definitions is not optimal to be used by libraries and LLMs.

MCP Version:

latest, 1.4.1

OS/Arc:

MacOS/ARM

[MacondoExpress]

@RobertoSannino
We fixed and issue that seems related to this one, can you confirm it has fixed also your issue so that we can close this as well?