CrewAI with an uncensored model

To run CrewAI with an uncensored model, build its LLM with custom_openai=True, a custom base URL and an abliterated model id.

Updated November 18, 2026By unrestrictedaiagent
crew.pyPython
import os
from crewai import Agent, Crew, LLM, Task

llm = LLM(
    model="kimi-k3.1-pro-uncensored",
    custom_openai=True,
    base_url="https://api.silkcompute.com",
    api_key=os.environ["SILK_API_KEY"],
)

analyst = Agent(
    role="Log analyst",
    goal="Find the errors in the logs the user provides",
    backstory="You read this team's own server logs.",
    llm=llm,
)
task = Task(
    description="Read {log} and list every error with its first timestamp.",
    expected_output="A table of errors and timestamps.",
    agent=analyst,
)
crew = Crew(agents=[analyst], tasks=[task])
print(crew.kickoff(inputs={"log": open("app.log").read()}))
OpenAI format
LLM(custom_openai=True)
Anthropic format
LLM(base_url=) on anthropic/
Tool calling
Agent(tools=)
Streaming
LLM(stream=True)
Browser use
Stagehand tool
Memory
Unified memory

Labels link to CrewAI's own docs, read November 18, 2026. Compare all 11

Point the agent only at systems, files and accounts you own or are allowed to test.

Deploy it

3 steps, with Kimi K3.1 Pro.

  1. 01

    Install CrewAI with the OpenAI extra. The native OpenAI client handles custom endpoints.

    Terminal
    pip install "crewai[openai]"
    export SILK_API_KEY=...
  2. 02

    Build the LLM. With custom_openai=True, CrewAI sends the model id as written to the base URL.

    See crew.py above
  3. 03

    Kick off the crew. Inputs fill the {braces} in each task.

    Terminal
    python crew.py

Which model

The config uses the Silk Compute base URL, which serves both models with a 1,048,576-token context.

Sources

More frameworks