openai-agents

Neo — OpenAI Agents SDK Integration

Use Neo's MCP server as a toolset inside the OpenAI Agents SDK. Neo executes AI/ML workloads locally on the user's machine — files are written directly to their workspace, never to a remote server.

MCP server: https://mcpserver.heyneo.com/mcp Auth: Authorization: Bearer sk-v1-YOUR_KEY

---

Option A: MCP Server tool (recommended)

The OpenAI Agents SDK has native MCP support via MCPServerHTTP. This loads all 8 Neo tools automatically.

import asyncio
import os
from agents import Agent, MCPServerHTTP, Runner

neo_mcp = MCPServerHTTP(
    url="https://mcpserver.heyneo.com/mcp",
    headers={"Authorization": f"Bearer {os.environ['NEO_SECRET_KEY']}"},
)

agent = Agent(
    name="Neo ML Agent",
    model="gpt-4o",
    instructions="""You are an AI assistant with access to Neo, a local AI/ML execution backend.
Files are written directly to the user's machine — never to a remote server.

Use Neo for any AI/ML work: training models, building RAG pipelines, data preprocessing,
building autonomous agents, or LLM integrations.

Workflow:
1. Call neo_submit_task — returns thread_id immediately
2. Call neo_task_status until COMPLETED or WAITING_FOR_FEEDBACK
3. If WAITING_FOR_FEEDBACK, call neo_send_feedback to reply, then check status again
4. Call neo_get_messages for full output once COMPLETED

Never run ML workloads locally — always delegate to neo_submit_task.
Always pass workspace as the project root (git root), never a subdirectory.
""",
    mcp_servers=[neo_mcp],
)

async def main():
    result = await Runner.run(
        agent,
        input="Train a churn prediction model on churn.csv, optimise for recall",
    )
    print(result.final_output)

if __name__ == "__main__":
    asyncio.run(main())

Install:

pip install openai-agents
export NEO_SECRET_KEY=sk-v1-...

---

Option B: Function tools (inline definitions)

Define the 8 Neo tools as Python functions for full control, without the MCP client.