Zenml

ZenML is a Munich, Germany-based company behind ZenML, an open-source framework for developing, evaluating and deploying classical machine learning models, LLM pipelines and AI agents. The core framework lets engineers write workflows ("pipelines") in Python that run on any infrastructure backend ("stacks"), and then operationalizes them by containerizing and tracking code, recording runs with metrics, logs and metadata, abstracting infrastructure, and integrating with existing tools such as MLflow, LangGraph, Langfuse, SageMaker and GCP Vertex. ZenML uses a client-server architecture with an integrated web dashboard, and can be installed locally or deployed as a remote server; it positions itself as a metadata layer over existing infrastructure so that data and compute remain on the customer's side.

The company also develops Kitaru, described as an open-source sister project focused on AI agents: "agent traces you can run, not just read." Kitaru records each agent run as a session with checkpoints for every model and tool call, either by wrapping an agent with an adapter or by importing existing traces (for example Langfuse traces or JSONL trace files). Recorded runs can be replayed against the team's real code with recorded tool calls answered from the recording, so replays do not touch live systems, and with one variable changed — a cheaper model, a new prompt, a tool policy, or a fix in the working tree. Sessions can be frozen into immutable cohorts, compared run-versus-run against a baseline with evaluators and cost metrics, and used to gate CI. Kitaru also provides durable-execution primitives (@flow, @checkpoint, wait(), resume, retry) for human-in-the-loop gates and for surviving crashes and long waits, with a separate lightweight self-hosted server and workers that execute replays in the user's environment. Adapters cover PydanticAI, the OpenAI Agents SDK, LangGraph, Mastra, the Vercel AI SDK and custom agent loops, and execution targets include inline execution, isolated jobs, sandboxes and external/MCP tools.

ZenML additionally maintains adjacent tooling: an MCP server (mcp-zenml) that lets users query pipelines, analyze runs and trigger deployments in natural language from MCP clients such as Claude Desktop or Cursor, a VS Code extension (ZenML Studio / vscode-zenml), agent "skills" repositories, and a dashboard/UI codebase.

Founding story

ZenML was founded in 2021 by Adam Probst and Hamza Tahir to address the difficulty of implementing and maintaining machine learning in production — vendor lock-in, high costs and tools that do not interoperate — by building an extensible, cloud- and tooling-agnostic open-source MLOps framework that expresses ML workflows as pipelines.

Business model

Open-source core distributed under Apache-2.0 with a commercial managed offering. ZenML and Kitaru can be self-hosted for free, while ZenML Pro (previously announced as ZenML Cloud) provides a managed control plane for running ML pipelines and agent flows with enterprise features such as single sign-on, role-based access control, audit logs, CI/CD integrations and dedicated support.

Paid managed/enterprise tier (ZenML Pro, launched as ZenML Cloud) layered on the free open-source frameworks, with features including SSO, RBAC, audit logs and dedicated support.

Traction

The zenml GitHub repository has about 5.6k stars, 654 forks and more than 8,900 commits; the kitaru repository has 268 stars and 17 forks. The ZenML organization hosts 85 public repositories, including a dashboard (63 stars), MCP server (49 stars) and VS Code extension (21 stars). The company reports usage by thousands of companies, citing Playtika and Leroy Merlin as users, and a Slack community of ML engineers.

Latest developments

The company is promoting Kitaru, its open-source agents project, across its homepage, documentation and GitHub organization. Kitaru covers replay-based regression testing (importing existing traces or capturing runs via adapters, freezing cohorts, running configuration experiments, gating CI) and durable agent execution (checkpoints, wait() for human approval, resume after crashes), with a separate lightweight self-hosted server, worker-based replay execution, a CLI, a UI repository and agent skills. Documentation is now organized into separate ZenML and Kitaru spaces, and repositories in the organization show commit activity through August 2026.