MLflow

Integrating Kern with MLflow

MLflow provides built-in GenAI tracing so you can capture, explore, and analyze LLM and agent traces. Kern integrates directly with MLflow via a single call to mlflow.kern.autolog().

Prerequisites

1. Install Dependencies

   Ensure the required packages are installed:

   1pip install -U mlflow kern-ai opentelemetry-exporter-otlp openinference-instrumentation-kern-ai

2. Start the MLflow tracking server

   Start the MLflow tracking server to view traces as you run your code:

   1mlflow server

   For more information on how to host an MLflow server, see the MLflow documentation.

   Tip

   If you don't want to self-host an MLflow server, you can use Managed MLflow offered by various cloud providers.

Set Environment Variables

Set the environment variables for the MLflow server URL and experiment name:

1export MLFLOW_TRACKING_URI="http://localhost:5000"
2export MLFLOW_EXPERIMENT_NAME="Kern Agent"

Alternatively, you can set these in your code using Python APIs. If you do this, you must call this before calling mlflow.kern.autolog().

1import mlflow
2
3mlflow.set_tracking_uri("http://localhost:5000")
4mlflow.set_experiment("Kern Agent")

Enable Automatic Tracing in Your Code

Call mlflow.kern.autolog() once at startup, then use your Kern agent as usual. MLflow will automatically record traces of model/tool calls and agent steps.

1import mlflow
2from kern.agent import Agent
3from kern.models.openai import OpenAIChat
4from kern.tools.yfinance import YFinanceTools
5
6# Enable MLflow tracing for Kern
7mlflow.kern.autolog()
8
9# Create and use the agent
10agent = Agent(
11    model=OpenAIChat(id="gpt-5-mini"),
12    tools=[YFinanceTools()],
13    instructions="Use tables to display data. Don't include any other text.",
14    markdown=True,
15)
16agent.print_response("What is the stock price of Apple?", stream=False)

View Traces

Access the MLflow UI to view the traces. If you started the UI locally, open http://127.0.0.1:5000 in your browser. If you are using a managed MLflow server, you can access the UI at the URL provided by the cloud provider.

AgentOS example

You can instrument your AgentOS application with MLflow by using the same approach as above. Simply call mlflow.kern.autolog() before creating your AgentOS instance.

1import mlflow
2from kern.agent import Agent
3from kern.db.sqlite import SqliteDb
4from kern.models.anthropic import Claude
5from kern.os import AgentOS
6from kern.tools.mcp import MCPTools
7
8# Setup automatic tracing for Kern
9mlflow.kern.autolog()
10
11agno_assist = Agent(
12    name="Kern Assist",
13    model=Claude(id="claude-sonnet-4-5"),
14    db=SqliteDb(db_file="kern.db"),                     # session storage
15    tools=[MCPTools(url="https://kern.ndx.rocks/mcp")],  # Kern docs via MCP
16    add_datetime_to_context=True,
17    add_history_to_context=True,                         # include past runs
18    num_history_runs=3,                                  # last 3 conversations
19    markdown=True,
20)
21
22# Serve via AgentOS → streaming, auth, session isolation, API endpoints
23agent_os = AgentOS(agents=[agno_assist], tracing=True)
24app = agent_os.get_app()

Then run your AgentOS application following the instructions. MLflow will automatically record traces of model/tool calls and agent steps.

Notes

• Ensure your model provider credentials (for example, OPENAI_API_KEY) are set in the environment.
• For best results, use the latest MLflow version that includes the Kern autolog integration.