zenml

ZenML is an open-source platform aimed at ML and AI engineers who work on classical machine learning, LLM workflows or agents in a company setting. Its core idea is that you write workflows (pipelines) in Python and run them on any infrastructure backend (stacks). Any Pythonic logic can be embedded in a pipeline, such as training a model or running an agentic loop. What the platform does around those pipelines: - Automatically containerizes and tracks your code. - Tracks individual runs with metrics, logs and metadata. - Abstracts away infrastructure complexity. - Integrates existing tools and infrastructure, with documented integrations for MLflow, LangGraph, Langfuse, SageMaker, GCP Vertex and others. - Provides an observable layer for iterating on experiments in development and production. Architecture and setup: ZenML uses a client-server architecture with an integrated web dashboard (a separate dashboard repository). A local mode runs client and server locally, while production deployments run the server separately and connect clients to it. Installation is via pip, with a server extra for full capabilities, followed by repository initialization and login. The README also mentions a VS Code / Cursor extension for managing pipelines from the editor. Documentation and examples: the project points to product documentation, a first-pipeline tutorial, a starter guide, an LLMOps guide and an SDK reference. Example projects in the repository cover a quickstart, agent architecture comparison, deploying ML models, deploying agents, end-to-end batch inference, an LLM RAG pipeline, an agentic deep-research workflow and a fine-tuning pipeline. Related tooling: an MCP server lets users query pipelines, analyze runs and trigger deployments through natural language from MCP-compatible clients. A sister project, Kitaru, focuses on replay-based evaluation for AI agents, with self-hosted execution and the stated split that ZenML is for ML pipelines while Kitaru is for agents. The README lists companies said to use ZenML, links to books featuring it, and describes contribution paths such as good-first-issues and writing integrations. It is distributed under the Apache License 2.0.