Skip to main content
CrewAI is an agent framework for orchestrating role-playing, autonomous AI agents. Agents are assigned roles, goals, and backstories, then collaborate to complete multi-step tasks. By routing CrewAI’s LLM calls through the Portkey AI Gateway, you gain visibility into cost, token usage, and latency across every agent run.

​
Installation

1

Start the gateway

npx @portkey-ai/gateway
The gateway listens at http://localhost:8787/v1.
2

Install dependencies

pip install crewai portkey-ai

​
Basic setup

CrewAI’s LLM class accepts base_url and extra_headers, making it straightforward to route through the gateway. 
from crewai import LLM, Agent, Task, Crew
from portkey_ai import createHeaders, PORTKEY_GATEWAY_URL

llm = LLM(
    model="gpt-4o",
    base_url=PORTKEY_GATEWAY_URL,
    api_key="dummy",  # Portkey handles auth via virtual key
    extra_headers=createHeaders(
        api_key="YOUR_PORTKEY_API_KEY",
        virtual_key="YOUR_VIRTUAL_KEY"
    )
)

agent = Agent(
    role="Researcher",
    goal="Research AI topics",
    backstory="An expert researcher with deep knowledge of AI systems.",
    llm=llm
)

task = Task(
    description="Summarize the key benefits of AI gateway routing.",
    expected_output="A concise bullet-point summary.",
    agent=agent
)

crew = Crew(agents=[agent], tasks=[task])
result = crew.kickoff()
print(result)
You can also use langchain_openai.ChatOpenAI with CrewAI if you prefer the LangChain integration. See the LangChain integration page for setup details.

​
Connecting via LangChain

If you are already using LangChain in your stack, pass a ChatOpenAI instance directly as the llm on each agent. 
from crewai import Agent, Task, Crew
from langchain_openai import ChatOpenAI
from portkey_ai import createHeaders, PORTKEY_GATEWAY_URL

llm = ChatOpenAI(
    model="gpt-4o",
    base_url=PORTKEY_GATEWAY_URL,
    api_key="dummy",
    default_headers=createHeaders(
        api_key="YOUR_PORTKEY_API_KEY",
        virtual_key="YOUR_VIRTUAL_KEY"
    )
)

agent = Agent(
    role="Researcher",
    goal="Research AI topics",
    backstory="An expert researcher with deep knowledge of AI systems.",
    llm=llm
)

​
Using multiple LLMs across agents

Assign different models to different agents to balance cost and capability. All requests remain visible in the Portkey observability dashboard. 
from crewai import Agent, Task, Crew, Process
from portkey_ai import createHeaders, PORTKEY_GATEWAY_URL
from crewai import LLM

# GPT-4o for strategic reasoning
gpt_llm = LLM(
    model="gpt-4o",
    base_url=PORTKEY_GATEWAY_URL,
    api_key="dummy",
    extra_headers=createHeaders(
        api_key="YOUR_PORTKEY_API_KEY",
        virtual_key="YOUR_OPENAI_VIRTUAL_KEY"
    )
)

# Claude Sonnet for writing tasks
anthropic_llm = LLM(
    model="claude-3-5-sonnet-latest",
    base_url=PORTKEY_GATEWAY_URL,
    api_key="dummy",
    extra_headers=createHeaders(
        api_key="YOUR_PORTKEY_API_KEY",
        virtual_key="YOUR_ANTHROPIC_VIRTUAL_KEY",
        config="YOUR_PORTKEY_CONFIG_ID",
        trace_id="crew-run-001"
    )
)

product_manager = Agent(
    role="Product Manager",
    goal="Define clear requirements for a software product",
    backstory="An experienced PM skilled in writing concise, developer-ready specs.",
    llm=gpt_llm
)

developer = Agent(
    role="Software Developer",
    goal="Develop software based on the provided requirements",
    backstory="A skilled developer proficient in building robust applications.",
    llm=anthropic_llm
)

task1 = Task(
    description="Define the key requirements for a classic ping-pong game.",
    expected_output="A clear, concise list of requirements.",
    agent=product_manager
)

task2 = Task(
    description="Based on the requirements, write Python code for the ping-pong game.",
    expected_output="Complete, runnable Python code.",
    agent=developer
)

crew = Crew(
    agents=[product_manager, developer],
    tasks=[task1, task2],
    process=Process.sequential,
    verbose=True
)

result = crew.kickoff()
print(result)

​
Real-world use case: web research crew

A two-agent crew that researches a topic and writes a report, with retry logic configured at the gateway. 
from crewai import Agent, Task, Crew, Process
from crewai import LLM
import json

config = {
    "retry": {"attempts": 5},
    "cache": {"mode": "simple"}
}

llm = LLM(
    model="gpt-4o-mini",
    base_url="http://localhost:8787/v1",
    api_key="sk-***",
    extra_headers={
        "x-portkey-config": json.dumps(config)
    }
)

researcher = Agent(
    role="Senior Research Analyst",
    goal="Uncover cutting-edge developments in AI",
    backstory="You work at a leading tech think tank. Your expertise lies in identifying emerging trends.",
    llm=llm
)

writer = Agent(
    role="Tech Content Strategist",
    goal="Craft compelling content on tech advancements",
    backstory="You are a renowned Content Strategist known for insightful articles on technology.",
    llm=llm
)

research_task = Task(
    description="Conduct a comprehensive analysis of the latest advancements in AI in 2025.",
    expected_output="Full analysis report in bullet points.",
    agent=researcher
)

write_task = Task(
    description="Compose an insightful article on the AI advancements from the research.",
    expected_output="A 4-paragraph article formatted as markdown.",
    agent=writer
)

crew = Crew(
    agents=[researcher, writer],
    tasks=[research_task, write_task],
    process=Process.sequential
)

result = crew.kickoff()
print(result)
The trace_id header groups all LLM calls from a single crew run in the Portkey dashboard, making it easy to debug individual runs and compare costs across executions.

Build docs developers (and LLMs) love

Get started for freeTalk to us