Connect CrewAI agents to Router One through one custom OpenAI-compatible endpoint

Install CrewAI and set your credentials

CrewAI requires Python 3.10 to 3.13. Install the crewai package in that environment; the OpenAI Python SDK this setup uses is a core dependency of crewai, so no crewai[litellm] extra is needed. The shell example below is for macOS/Linux. Replace both placeholders with a Router One key and the exact ID of a current model that supports Chat Completions, keeping any prefix that is part of the ID (for example anthropic/). The ROUTER_ONE_* names belong to this example and are read explicitly by the Python file; keep the same environment active when running it. The behavior described on this page was checked against crewai 1.15.20.

python -m pip install crewai
export ROUTER_ONE_API_KEY="sk-your-router-one-key"
export ROUTER_ONE_MODEL_ID="<exact-model-id-from-/models>"

Configure CrewAI to use the Router One base URL

Save this as crewai_router_one.py and run python crewai_router_one.py. LLM(custom_openai=True, base_url=..., api_key=...) is the mode CrewAI documents for a custom OpenAI-compatible endpoint: it always uses the OpenAI SDK's Chat Completions path, POST /v1/chat/completions, whatever the model ID starts with. The model string is openai/ followed by the exact catalog ID. In this mode CrewAI removes exactly one leading openai/ segment and sends the remainder unchanged, so anthropic/claude-sonnet-5 or openai/gpt-5.5 reaches the gateway exactly as the catalog lists it. The agent has no tools and the task expects plain text, so the first request carries only model and messages. crew.kickoff() runs the task and returns a CrewOutput; result.raw is the model's text.

import os

from crewai import Agent, Crew, LLM, Task

llm = LLM(
    model="openai/" + os.environ["ROUTER_ONE_MODEL_ID"],
    custom_openai=True,
    base_url="https://api.router.one/v1",
    api_key=os.environ["ROUTER_ONE_API_KEY"],
)
greeter = Agent(
    role="Greeter",
    goal="Reply with one short greeting.",
    backstory="You keep answers short and literal.",
    llm=llm,
)
task = Task(
    description="Reply with one short greeting.",
    expected_output="One short greeting sentence.",
    agent=greeter,
)
crew = Crew(agents=[greeter], tasks=[task])
result = crew.kickoff()
print(result.raw)

Use environment variables instead of an LLM object

A project scaffolded with crewai create crew builds its agents from config/agents.yaml without an llm field, and CrewAI then reads the model and endpoint from the environment: MODEL selects the model, OPENAI_API_BASE (OPENAI_BASE_URL is also accepted) sets the endpoint, and OPENAI_API_KEY supplies the key. CrewAI calls load_dotenv() when it imports its LLM module; exporting the variables in the shell works in every project layout. Use the same openai/ plus catalog ID form for MODEL: when the base URL comes from these variables, CrewAI attaches it while building the LLM, recognizes that the ID is not an official OpenAI name, and takes the same custom endpoint path, so the exact ID is sent. A plain string in agents.yaml (llm: openai/anthropic/claude-sonnet-5) or Agent(llm="...") is built without that base URL and is handed to CrewAI's LiteLLM fallback instead, which is not installed by default; keep the model in MODEL or in an LLM object.

# .env in the project directory, or export the same names in your shell
MODEL=openai/<exact-model-id-from-/models>
OPENAI_API_BASE=https://api.router.one/v1
OPENAI_API_KEY=sk-your-router-one-key

What to check before adding tools, structured output, or more agents

Agent tools, including the delegation tools added when allow_delegation=True and anything from the crewai-tools package, are sent as function-tool schemas on the same Chat Completions request; check tool calling on the model's detail page before enabling them. Task output_pydantic or output_json and LLM(response_format=...) send a JSON-schema response_format, so check structured-output support for the exact model first. Each Agent can take its own LLM object, so one crew can mix models on one key, and every call still lands in the same trace. LLM(stream=True) keeps the same path and adds stream_options so usage arrives in the final chunk. Leave api at its default: api="responses" switches the same object to POST /v1/responses, which the gateway serves natively only for the GPT-family IDs whose detail page lists that endpoint. A crew with several agents, tasks, or retries makes several model requests, and crew.usage_metrics is CrewAI's own token tally, not a bill. Give the crew a dedicated key with maxSpend, then reconcile in Dashboard → Logs by key, time, exact model, and request_id.

Which model ID should CrewAI send?

Copy the exact model ID from /models, preserving case, hyphens, and version suffixes; do not substitute a display name. Open its detail page and match the supported API endpoints, context window, and capabilities such as tool calling to the provider and features selected in CrewAI. A catalog listing does not mean the client can use every feature of that model. Give each tool a dedicated API key with a maxSpend cap.

Which API protocol is CrewAI using?

OpenAI-compatible describes an interface format; it does not make Chat Completions (/v1/chat/completions), Responses (/v1/responses), and Anthropic Messages (/v1/messages) interchangeable. Check the installed client version, provider configuration, and actual request path against the model detail page and API compatibility fact sheet. A successful plain-text chat does not establish support for hosted tools, conversation state, or file-editing features.

Verify the CrewAI call in your request trace

Send a simple text request from CrewAI, then match its trace in Dashboard → Logs by time, model, and request_id: tokens, cost, latency, and status. Next, test streaming, tool calls, and multi-turn history separately. For failures, retain the actual request path, full error message, and request_id. If there is no matching log, check client configuration and connectivity before attributing the error to the gateway or upstream.

FAQ