ZenML

ZenML is a Python framework for ML and AI engineers who run classical ML training, LLM workflows, or agents in a company setting. You write workflows as pipelines of steps containing ordinary Python (a training loop, an agentic loop) and run them on a configurable infrastructure backend called a stack. ZenML then containerizes and tracks the code, records each run with metrics, logs, and metadata, and hides the details of the underlying orchestrator and storage.

It uses a client-server architecture with a web dashboard. A local install runs client and server on one machine; in production the server is deployed separately and clients connect with zenml login. Existing code is wrapped in a @step decorator rather than rewritten, so scikit-learn, PyTorch, LangGraph, LlamaIndex, or raw API calls keep working, and integrations connect tools you already run: MLflow and Weights & Biases for experiment tracking, Langfuse, SageMaker, Vertex AI, and Kubernetes-based orchestrators including the native Kubernetes orchestrator and Kubeflow.

Around the core are a ZenML MCP server that lets Claude Desktop, Cursor, or another MCP client query pipeline runs, inspect metrics, and trigger deployments in natural language; a VS Code and Cursor extension for managing pipelines from the editor; and a managed ZenML Pro control plane. A sister project, Kitaru, handles agent evaluation by recording agent runs (a PydanticAI wrapper or imported Langfuse traces) and replaying them against your real code with recorded tool calls, so you can fork a run with a cheaper model or a new prompt and compare results, group traffic into cohorts, and gate CI on the outcome.

Features

• Pipelines and steps: Python functions decorated as steps compose into versioned, reproducible workflows
• Stacks: the same pipeline runs on local, Kubernetes, Kubeflow, SageMaker, or Vertex backends by switching stack configuration
• Containerization and tracking: code is packaged automatically and every run stores artifacts, metrics, logs, and metadata
• Deployments and snapshots: pipelines deploy models and agent services as endpoints with monitoring and versioning
• Tool integrations: MLflow, Weights & Biases, Langfuse, LangGraph, LiteLLM, and cloud orchestrators
• Dashboard: a web UI for runs, artifacts, and models, with a client-server split for production
• MCP server: natural-language queries over pipeline metadata and run triggering from MCP clients
• Editor extension: pipeline management from VS Code and Cursor
• Kitaru: replay-based agent evals with cohorts, experiments, and CI gating, self-hosted on your infrastructure
• Examples: agent architecture comparison, ML model deployment, agent deployment, end-to-end batch inference, RAG, deep research, and fine-tuning pipelines