CrewAI with EXAIV: Pointing a Multi-Agent Crew at OpenAI-Compatible Domestic Models

CrewAI is a Python framework for orchestrating multiple agents that divide a job between them. It connects to models through native SDK integrations for the largest providers and falls back to LiteLLM for everything else — and LiteLLM accepts any OpenAI-compatible endpoint, which is the door EXAIV comes through.

Two things make this integration worth reading before you copy code. First, CrewAI has a default model (gpt-4o-mini unless you configure otherwise), so a partially configured project will try to reach OpenAI even when you thought you pointed it elsewhere. Second, because each agent takes its own llm argument, a crew can mix a cheap fast model with a stronger one — which is the reason to run a crew through a relay in the first place.

Setup steps

Install CrewAI with the LiteLLM extra
uv add 'crewai[litellm]' (or the pip equivalent). The docs note LiteLLM is the fallback for providers without a native integration.
Create an EXAIV API key
Generate one in the EXAIV console; the key identifies your account to the relay.
Configure the endpoint
Pick one of the two routes below — LLM class or environment variables. Do not mix them halfway; a leftover OPENAI_MODEL_NAME from a tutorial is a common cause of confusing failures.
Verify with a single agent before building the crew
Run one agent with one task and print the result. Debug an orchestrator and an endpoint at the same time only if you enjoy guesswork.

Route 1: the LLM class

The explicit route keeps configuration in code, which is easier to review and to keep per-environment.

import os
from crewai import Agent, Task, Crew, Process, LLM

llm = LLM(
    model="openai/kimi-k3",                       # openai/ prefix = the OpenAI-compatible path
    base_url="https://exaiv.com/api/ai/v1",
    api_key=os.environ["EXAIV_API_KEY"],
)

researcher = Agent(
    role="Researcher",
    goal="Collect the facts needed for the brief",
    backstory="You gather sources and never invent a citation.",
    llm=llm,
)

writer = Agent(
    role="Writer",
    goal="Turn the research notes into a readable brief",
    backstory="You write plainly and avoid unverifiable claims.",
    llm=LLM(
        model="openai/deepseek-v4.1-flash",
        base_url="https://exaiv.com/api/ai/v1",
        api_key=os.environ["EXAIV_API_KEY"],
    ),
)

gather = Task(
    description="List the facts and open questions for: {topic}",
    expected_output="A bulleted fact list with sources named for each item.",
    agent=researcher,
)

write = Task(
    description="Write a one-page brief for: {topic}",
    expected_output="A structured brief with a heading and three sections.",
    agent=writer,
)

crew = Crew(
    agents=[researcher, writer],
    tasks=[gather, write],
    process=Process.sequential,
    verbose=True,
)

print(crew.kickoff(inputs={"topic": "open-weight model releases"}))

Route 2: environment variables

The environment route configures the whole process at once, which suits container deployments and CI runs.

import os
from crewai import Agent, Task, Crew, Process

os.environ["OPENAI_API_KEY"] = os.environ["EXAIV_API_KEY"]
os.environ["OPENAI_API_BASE"] = "https://exaiv.com/api/ai/v1"
os.environ["OPENAI_MODEL_NAME"] = "openai/glm-5.3"

agent = Agent(
    role="Analyst",
    goal="Explain the trade-offs in the provided options",
    backstory="You state assumptions explicitly.",
)

task = Task(
    description="Compare the options in {notes} and recommend one.",
    expected_output="A short comparison plus a single recommendation.",
    agent=agent,
)

crew = Crew(agents=[agent], tasks=[task], process=Process.sequential)
print(crew.kickoff(inputs={"notes": "option A: faster, option B: cheaper"}))

Assigning models per agent

Because every agent accepts its own llm, the crew's shape determines which models you need. A practical split for a research-and-write crew:

All model identifiers must come from the live catalog on the EXAIV side (/models); a name memorised from an upstream provider will not resolve through the relay.

Running the same crew against a local server

The confusing part of multi-agent frameworks is usually the framework, not the model. That is an argument for testing the orchestration logic against a local OpenAI-compatible server first, then switching the endpoint:

from crewai import LLM

local = LLM(
    model="ollama/llama3.2",                    # documented local-model form
    base_url="http://localhost:11434",          # Ollama default port
)

# switching to the relay changes only these two lines: model and base_url
relay = LLM(
    model="openai/deepseek-v4.1-flash",
    base_url="https://exaiv.com/api/ai/v1",
    api_key="YOUR_EXAIV_KEY",
)

This is also the practical answer for privacy-sensitive runs: the crew logic stays identical, and only the endpoint changes when the data cannot leave your own hardware.

Troubleshooting

The recurring lesson in this table: most "endpoint problems" in a crew are configuration problems in the crew. Verify with a single agent before debugging the orchestration.

Verification checklist

• One agent with one task returns a real answer through EXAIV
• The model identifier matches the live catalog
• No leftover OPENAI_* variable points at a different provider
• Each agent's llm is set explicitly, including the ones you forgot
• A failure in the middle of a crew produces a readable error, not a silent empty result
• If the crew runs in a container, all three configuration variables are injected

Related reading

• Chat Completions reference
• Model catalog
• Pricing and credits
• Agent framework guide
• Local deployment playbook