OpenAI Agents SDK

Give your OpenAI Agents SDK agents intelligent access to 150+ integrations using Pica’s MCP Server. The Model Context Protocol (MCP) integration provides seamless tool access with built-in authentication, error handling, and Pica’s knowledge base.

For complete MCP Server documentation including all tools and configuration options, see the MCP Server guide. This page focuses on OpenAI Agents SDK-specific usage patterns.

Prerequisites

Pica Account - Free account for managing integrations
Python 3.10+ - Required for OpenAI Agents SDK
OpenAI Agents SDK - Installed via pip

Installation

pip install openai-agents

Quick Start

Create an agent with Pica’s MCP Server:

from agents import Agent, Runner
from agents.extensions.mcp import MCPServerManager
import os

# Initialize MCP manager and add Pica server
mcp = MCPServerManager()
mcp.add_server(
    name="pica",
    command="npx",
    args=["@picahq/mcp"],
    env={"PICA_SECRET": os.getenv("PICA_SECRET")}
)

# Create agent with MCP tools
agent = Agent(
    name="integration-agent",
    instructions="""You are an AI assistant with access to integrations through Pica.
    You can list available integrations, search for actions, get detailed action knowledge,
    and execute actions on behalf of the user.""",
    mcp_servers=[mcp]
)

# Run the agent
result = Runner.run_sync(
    agent,
    "What integrations do I have connected?"
)

print(result.final_output)

Usage Patterns

Basic Agent with MCP Server

The simplest setup - an agent with access to all MCP tools:

Example

from agents import Agent, Runner
from agents.extensions.mcp import MCPServerManager
import os

# Setup MCP server
mcp = MCPServerManager()
mcp.add_server(
    name="pica",
    command="npx",
    args=["@picahq/mcp"],
    env={"PICA_SECRET": os.getenv("PICA_SECRET")}
)

# Create agent
agent = Agent(
    name="automation-agent",
    instructions="""You help users automate tasks across their integrated tools.
    Use Pica's tools to discover available integrations, search for actions,
    understand action requirements, and execute actions.""",
    model="gpt-5",
    mcp_servers=[mcp]
)

# Execute tasks
result = Runner.run_sync(
    agent,
    "Send an email via Gmail to team@company.com with the subject 'Weekly Update'"
)

print(result.final_output)

Agent with Handoffs

Use multiple specialized agents with handoff capabilities. This example combines a Pica-powered integration agent with a text analysis agent:

Example

from agents import Agent, Runner, Handoff
from agents.extensions.mcp import MCPServerManager
import os

# Setup MCP server
mcp = MCPServerManager()
mcp.add_server(
    name="pica",
    command="npx",
    args=["@picahq/mcp"],
    env={"PICA_SECRET": os.getenv("PICA_SECRET")}
)

# Integration specialist (with Pica MCP)
integration_agent = Agent(
    name="integration-agent",
    instructions="""You specialize in working with third-party integrations.
    You can list integrations, search for actions, and execute operations
    through Gmail, Slack, Salesforce, and other platforms.""",
    model="gpt-5",
    mcp_servers=[mcp]
)

# Text analysis specialist (no Pica tools)
analyst_agent = Agent(
    name="analyst-agent",
    instructions="""You specialize in analyzing and summarizing text content.
    You don't have access to integrations, but you're excellent at
    understanding patterns, extracting insights, and creating summaries.""",
    model="gpt-5"
    # Note: No mcp_servers - this agent doesn't use Pica
)

# Coordinator agent
coordinator = Agent(
    name="coordinator",
    instructions="""You coordinate between integration operations and text analysis.
    
    Hand off to:
    - integration-agent: For fetching data from or sending data to integrations
    - analyst-agent: For analyzing, summarizing, or processing text content
    
    Example workflow: Have integration-agent fetch emails, then have analyst-agent
    summarize them, then have integration-agent send the summary.""",
    model="gpt-5",
    handoffs=[
        Handoff(agent=integration_agent, description="Work with integrations"),
        Handoff(agent=analyst_agent, description="Analyze and summarize text")
    ]
)

# Run with handoffs
result = Runner.run_sync(
    coordinator,
    """Get my last 5 Gmail emails, analyze their sentiment and key topics,
    then send a summary to team@company.com"""
)

print(result.final_output)

Agent with Guardrails

Add validation and safety checks to your agent. This example blocks emails to restricted addresses:

Example

from agents import Agent, Runner, Guardrail
from agents.extensions.mcp import MCPServerManager
import os
import re

# Setup MCP server
mcp = MCPServerManager()
mcp.add_server(
    name="pica",
    command="npx",
    args=["@picahq/mcp"],
    env={"PICA_SECRET": os.getenv("PICA_SECRET")}
)

# Define email validation guardrail
def validate_email_recipient(context):
    """Block emails to restricted addresses"""
    message = context.input.lower()
    
    # Extract email addresses from the request
    email_pattern = r'\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b'
    emails = re.findall(email_pattern, message)
    
    # Blocked email addresses
    blocked_emails = ["john@example.com", "blocked@company.com"]
    
    for email in emails:
        if email.lower() in blocked_emails:
            return False, f"Cannot send email to {email} - this address is restricted"
    
    return True, None

# Create email agent with guardrails
email_agent = Agent(
    name="email-agent",
    instructions="""You help users send emails using Gmail.
    Search for Gmail email sending actions and execute them with the user's parameters.""",
    model="gpt-5",
    mcp_servers=[mcp],
    guardrails=[
        Guardrail(name="email-validator", fn=validate_email_recipient, on="input")
    ]
)

# This will work
try:
    result1 = Runner.run_sync(
        email_agent,
        "Send an email to team@company.com with subject 'Meeting Update'"
    )
    print("✅ Email sent:", result1.final_output)
except Exception as e:
    print(f"❌ Blocked: {e}")

# This will be blocked by guardrail
try:
    result2 = Runner.run_sync(
        email_agent,
        "Send an email to john@example.com saying hello"
    )
    print("✅ Email sent:", result2.final_output)
except Exception as e:
    print(f"❌ Blocked: {e}")

Streaming Responses

Stream agent responses in real-time:

Example

from agents import Agent, Runner
from agents.extensions.mcp import MCPServerManager
import os

# Setup MCP server
mcp = MCPServerManager()
mcp.add_server(
    name="pica",
    command="npx",
    args=["@picahq/mcp"],
    env={"PICA_SECRET": os.getenv("PICA_SECRET")}
)

# Create agent
agent = Agent(
    name="streaming-agent",
    instructions="""You help users discover and work with their Pica integrations.
    Explain what you're doing as you work through the process.""",
    model="gpt-5",
    mcp_servers=[mcp]
)

# Stream the response
async def stream_agent():
    async with Runner.run_stream(
        agent,
        "What integrations do I have connected in Pica?"
    ) as stream:
        async for event in stream:
            if event.type == "agent_response":
                print(event.data.content, end="", flush=True)
            elif event.type == "tool_call":
                print(f"\n[Using tool: {event.data.name}]")
            elif event.type == "tool_result":
                print(f"[Tool completed]")

# Run async
import asyncio
asyncio.run(stream_agent())

Multi-Session Support

Use sessions to maintain conversation history:

Example

from agents import Agent, Runner, Session
from agents.extensions.mcp import MCPServerManager
import os

# Setup MCP server
mcp = MCPServerManager()
mcp.add_server(
    name="pica",
    command="npx",
    args=["@picahq/mcp"],
    env={"PICA_SECRET": os.getenv("PICA_SECRET")}
)

# Create agent
agent = Agent(
    name="context-agent",
    instructions="You help with integrations and remember context from previous messages.",
    model="gpt-5",
    mcp_servers=[mcp]
)

# Create session
session = Session()

# First interaction
result1 = Runner.run_sync(
    agent,
    "What integrations do I have?",
    session=session
)

# Agent remembers previous context
result2 = Runner.run_sync(
    agent,
    "Show me actions for the first one",
    session=session
)

print(result2.final_output)

Configuration

The MCP server accepts configuration through environment variables:

mcp.add_server(
    name="pica",
    command="npx",
    args=["@picahq/mcp"],
    env={
        "PICA_SECRET": os.getenv("PICA_SECRET"),  # Required: Your Pica API key
        # Optional: Add custom headers
        "CUSTOM_HEADER": "value"
    }
)

Full MCP configuration options

View all environment variables and configuration settings

Best Practices

Use specific instructions

Provide clear instructions about when and how to use Pica tools:

instructions = """You are an integration assistant with access to Pica tools.

Always follow this workflow:
1. Use list_pica_integrations to see available platforms
2. Use get_pica_platform_actions to find relevant actions
3. Use get_pica_action_knowledge to understand requirements
4. Use execute_pica_action to perform the operation

Ask for clarification if you're unsure about any parameters."""

Implement guardrails

Always add guardrails for production agents to prevent:

Destructive operations without confirmation
Exposure of sensitive information
Invalid or malicious requests

Use handoffs for complex workflows

Split complex tasks across specialized agents:

Data retrieval agent (read-only)
Action execution agent (write operations)
Coordinator agent (orchestrates handoffs)

Leverage sessions

Use sessions to maintain context across multiple interactions, making your agent feel more natural and contextual.

Enable tracing

Use OpenAI Agents SDK’s built-in tracing to debug and monitor your agent:

from agents import trace

# Enable tracing
trace.configure(enabled=True)

# Run agent with traces
result = Runner.run_sync(agent, "task")

Troubleshooting

MCP server not starting

Problem: Agent can’t connect to MCP server.Solutions:

Ensure Node.js is installed: node --version
Verify PICA_SECRET is set: echo $PICA_SECRET
Check MCP server logs for errors
Try running manually: npx @picahq/mcp

Tools not available

Problem: Agent says Pica tools are not available.Solutions:

Verify MCP server is added: mcp_servers=[mcp] in Agent
Check that PICA_SECRET is valid
Ensure you have connections in Pica dashboard

Action execution failing

Problem: execute_pica_action returns errors.Solutions:

Always call get_pica_action_knowledge first
Verify connection key is correct
Check required parameters are provided
Ensure connection is authenticated

Agent context too large

Problem: Token limit exceeded.Solutions:

Use more specific queries
Limit number of actions searched
Use streaming for long responses
Consider using handoffs to specialized agents

What’s next?

MCP Server Documentation

Complete guide to Pica’s MCP Server tools and configuration