How much does ZenML cost?

Open source ZenML and Kitaru are free under Apache 2.0. ZenML Pro Starter is $39 a month for one AI agent, one ML project, up to 100 pipeline runs a month and unlimited seats, with a 14 day trial and no card. Enterprise is quoted and adds more agents and projects, on premises or air gapped installs, the customer's own LLM endpoint, SSO, SCIM, custom roles and audit logs. Replays and model judges run on the customer's own model keys.

Does ZenML support MCP and tool calling?

Full under the Agentic Index verification standard. ZenML publishes public evidence of Model Context Protocol (MCP) support, tool calling, and API or SDK extensibility for connecting agents to external tools. Evidence basis: ZenML has a Python SDK, a CLI and a REST API, with a separate ZenML Pro API covering organizations, roles, role assignments and permission checks, and Bots chats and artifacts can be created through the API as well. Kitaru offers Python and. Graded from public evidence only against the Agentic Index 14 point capability taxonomy. No vendor pays for placement.

Does ZenML have SOC 2, SSO and role based access control?

Full under the Agentic Index verification standard. ZenML publishes public evidence of enterprise security and identity controls, covering a compliance posture such as SOC 2 or ISO 27001, single sign on (SSO or SAML), and role based access control (RBAC). Evidence basis: ZenML Pro controls access through roles with customizable permissions, teams and service accounts, and each bot works through its own service account with a Bot role that decides which workspaces and resources it can read, narrowed further. Graded from public evidence only against the Agentic Index 14 point capability taxonomy. No vendor pays for placement.

Can ZenML be self hosted or run in a private VPC?

Full under the Agentic Index verification standard. ZenML publishes public evidence of self hosted or private VPC deployment, along with data residency options for controlling where agent data is processed and stored. Evidence basis: Open source ZenML and Kitaru run wherever the customer installs them, from a laptop to Docker or Kubernetes with Helm against the customer's own Postgres, with workers running replays and evaluators inside the customer's environment. ZenML. Graded from public evidence only against the Agentic Index 14 point capability taxonomy. No vendor pays for placement.

Back to vendors

ZenML

Q: What are the best alternatives to ZenML?

Based on documented capability coverage across the Agentic Index 14 point taxonomy, the closest alternatives to ZenML among agent infrastructure platforms are Windmill, PostHog, Agno, Haystack. Each is graded on the same public evidence, and no vendor pays for placement.

Also known as: ZenML Labs, ZenML GmbH, zenml.io, ZenML Pro, Kitaru

Visit site

Agent infrastructureindependentVerified 2026-11-09

Entry priceOpen source ZenML and Kitaru are free. ZenML Pro Starter is $39 a month with unlimited seats; Enterprise is quoted.Full pricing detail

ZenML Labs, the Munich company behind the ZenML pipeline framework, sells tools for teams running AI agents and ML pipelines in production. Kitaru, open source, imports agent traces from Langfuse, LangSmith and other tools, finds patterns, turns them into evaluators and replays a fix against recorded sessions. ZenML orchestrates pipelines and agents on Kubernetes, Vertex AI or SageMaker. ZenML Pro, from $39 a month, hosts both and adds AI engineers: coding agents that watch traces and runs, test fixes and open pull requests.

ZenML is made by ZenML GmbH of Munich, Germany, a remote team with an office in Munich that began with the open source ZenML framework for machine learning pipelines and now sells two open source products and a hosted platform for teams running AI agents in production. Both products carry the Apache 2.0 license. Customer stories include Cross Screen Media, which trains models for 210 markets with ZenML, and Zuiver.ai's AI detection pipeline, and the company keeps a searchable database of more than 2,200 production LLMOps case studies alongside its own agent benchmarks on claims, appeals and underwriting work.

Kitaru works on the agent a team already runs. It imports traces from Langfuse, LangSmith, Braintrust, Logfire, MLflow and Arize Phoenix, or records new runs through adapters for PydanticAI, LangGraph, the OpenAI Agents SDK, the Claude Agent SDK, Mastra and the Vercel AI SDK. Each run becomes a session holding every model call, tool call and decision with tokens and cost.

Analyzers surface patterns across sessions, reviewers pin their judgments to exact traces in investigations, findings become evaluators written as code checks or typed model judges, and a replay reruns failed sessions with a swapped model, prompt or code branch, with tool policies deciding whether tool calls use recorded results or live tools. A coding assistant drives the loop through Kitaru's MCP server and skills.

ZenML itself orchestrates Python pipelines and agents on Kubernetes, Vertex AI, SageMaker or AzureML, versions every artifact and model, and serves a pipeline as an HTTP deployment. ZenML Pro adds Bots: coding agents that work in a sandbox per chat with the team's repositories, reach GitHub, Notion, Context7, Firecrawl, Tavily or a custom MCP server through connectors, take tasks from the dashboard, Slack or GitHub, run on cron automations and open pull requests with fixes. The hosted AI engineers built on them, an eval engineer, a reliability engineer and a cost engineer, read traces and pipeline runs and test each fix against what runs today.

Open source ZenML and Kitaru are free to self host. ZenML Pro Starter costs $39 a month for one AI agent and one ML project with up to 100 pipeline runs a month and unlimited seats, after a 14 day trial with no card. Enterprise is quoted and adds more agents and projects, multiple workspaces, on premises or air gapped installs, the customer's own LLM endpoint, SAML or OIDC single sign on, SCIM, custom roles, audit logs and an SLA.

ZenML Pro carries role based access with teams and service accounts, and the company holds SOC 2 Type II and ISO 27001 certification. On Starter the AI engineers run in ZenML's cloud and read traces through revocable keys while data, artifacts and compute stay in the customer's cloud; self hosted and on premises installs keep everything inside the customer's infrastructure.

ZenML fits platform and ML teams that already run agents on common frameworks and want trace review, regression tests and model swap experiments without leaving their tracing tool, and teams that orchestrate training, fine tuning and batch inference across clouds. It fits less well for buyers who want a hosted agent builder or a managed knowledge base for their own agents, a choice of the model behind the hosted AI engineers on the Starter plan, or an agent that operates a browser.

Vendor details

Canonical URL

https://www.zenml.io

Subcategory

Agent trace evaluation and replay, pipeline orchestration and hosted AI engineer agents, open source with a paid platform

Funding status

Privately held. ZenML GmbH is registered in Munich; no funding rounds are named on the pages read.

Company status

independent

Use cases & customers

Primary use cases

importing agent traces and finding repeated failure patterns across every sessionturning reviewer judgments into code checks and typed model judges that run on every sessionreplaying failed sessions with a swapped model, prompt or code branch before shipping a changeorchestrating Python pipelines and agents on the customer's own orchestrator and cloudhanding pipeline failures and agent regressions to hosted coding agents that open pull requests

Target customers

ML and platform teams running AI agents in production on PydanticAI, LangGraph, the OpenAI Agents SDK, the Claude Agent SDK or Mastrateams already tracing agents in Langfuse, LangSmith, Braintrust or Arize Phoenix that want evaluations and replays on topML engineering teams orchestrating training, fine tuning and batch inference across Kubernetes, Vertex AI, SageMaker or AzureMLinsurers and financial services teams that need on premises or air gapped deployment

Deployment options

Open source ZenML and Kitaru self hosted on a laptop, Docker or Kubernetes with HelmZenML Pro SaaSZenML Pro hybrid, with the control plane hosted by ZenML and workspaces in the customer's infrastructureZenML Pro self hosted, on premises or air gapped on Enterprise

Integrations

Kitaru imports traces from Langfuse, LangSmith, Braintrust, Logfire, MLflow and Arize Phoenix, records runs through adapters for PydanticAI, LangGraph, the OpenAI Agents SDK, the Claude Agent SDK, Mastra and the Vercel AI SDK, and works from Claude Code, Codex and Cursor through its MCP server and skills. ZenML runs pipelines on Kubernetes, Vertex AI, SageMaker and AzureML with service connectors for cloud access. ZenML Pro Bots connect to GitHub, Slack, Notion, Context7, Firecrawl, Tavily and custom MCP servers.

In practice

A support team imports 1,824 Langfuse traces from its returns agent into Kitaru and asks its coding assistant to investigate. The assistant reads twenty sessions and finds that the agent drops a hazardous materials flag when customers mention batteries, and a reviewer confirms it in the UI. Kitaru freezes the 90 affected sessions as a cohort, drafts an evaluator and replays them against a pull request branch, where failures fall from 90 to 4.

An ML team sets a ZenML Pro automation that starts a chat every morning with a prompt to check failed pipeline runs. The bot opens a sandbox with the team's repository, inspects last night's failed training run in the ZenML workspace, fixes the step and opens a draft pull request, then asks for approval before it reruns the pipeline.

An insurer weighing a cheaper model for its claims agent replays the same cohort of 200 recorded sessions with only the model swapped. Kitaru compares the two runs on the team's evaluators and shows cost per session down 61 percent with 196 of 200 decisions unchanged, and the same experiment then runs on every commit so the regression cannot return unseen.

Sources & related URLs

Related / legacy domains

https://docs.zenml.io https://docs.zenml.io/kitaru https://www.zenml.io/pricing https://security.zenml.io https://github.com/zenml-io

Research sources

https://www.zenml.io https://www.zenml.io/product/kitaru https://www.zenml.io/product/zenml https://www.zenml.io/pricing https://www.zenml.io/imprint https://www.zenml.io/company https://docs.zenml.io/pro/bots/bots https://docs.zenml.io/pro/bots/bots/bots-core-concepts https://docs.zenml.io/pro/bots/bots/bots-system-architecture https://docs.zenml.io/pro/core-concepts/triggers https://docs.zenml.io/pro/readme https://docs.zenml.io/kitaru/getting-started/setup https://docs.zenml.io/kitaru/guides/replay-and-overrides https://docs.zenml.io/kitaru/guides/judge-evaluations

Agentic Index coverage score

12.5 / 14 capabilities · 89%

Integrations & Tool Calling	Full
Agents running in ZenML Pro reach outside systems through connectors set up once for the organization: GitHub, Notion, Context7, Firecrawl and Tavily, or any custom MCP server started inside the sandbox. Through the ZenML Pro GitHub app an agent clones the team's repositories, pushes branches and opens pull requests with its fixes, and an egress proxy adds each credential on the way out so the sandbox never holds it. Pipelines reach AWS, Google Cloud, Azure and Kubernetes through service connectors. Sourcedocs.zenml.io/pro/bots/botsread 2026-11-09
Workflow Orchestration	Full
ZenML pipelines chain Python steps into a graph and run it on the orchestrator a team already has, Kubernetes, Vertex AI, SageMaker or AzureML, with every artifact and model version tracked between steps, and the same pipeline can be served as an HTTP deployment for real time calls. Agents run as pipeline steps or deployments. In ZenML Pro a prompt goes to the bot whose description fits it best, and within one chat a bot inspects pipeline runs, changes code, reruns a pipeline once approved and reports back. Sourcezenml.io/product/zenmlread 2026-11-09
Knowledge Grounding & RAG	Partial
Agents pull context in at run time rather than from a knowledge store ZenML maintains for them. Notion pages, current library documentation through Context7 and web search results through Tavily reach a bot through connectors when a task calls for them, and a chat can point the agent at specific pipeline runs, model versions, Kitaru sessions or files. ZenML keeps no managed index or embeddings layer over a customer's own documents that agents search across runs, so retrieval for the customer's own agents stays in the customer's code. Sourcedocs.zenml.io/pro/bots/bots/bots-core-conceptsread 2026-11-09
Human Oversight & Guardrails	Full
A ZenML Pro bot asks for approval before it reruns, replays or promotes anything the person did not ask for, and when it needs a decision it puts a question to the chat, often with options and a recommended choice, and waits in a blocked state until someone answers. People can send new instructions mid turn, queue them, interrupt the agent or open an SSH session into its sandbox. Opening a draft pull request needs no extra approval, so code changes still pass the team's own review. Sourcedocs.zenml.io/pro/bots/botsread 2026-11-09
Security, Identity & Governance	Full
ZenML Pro controls access through roles with customizable permissions, teams and service accounts, and each bot works through its own service account with a Bot role that decides which workspaces and resources it can read, narrowed further with custom roles. Enterprise adds SAML or OIDC single sign on, SCIM and audit logs. The company holds SOC 2 Type II and ISO 27001 certification and runs a trust center for security reviews. Credentials stay outside agent sandboxes, swapped in by an egress proxy only for approved destinations. Sourcedocs.zenml.io/pro/readmeread 2026-11-09
Observability & Auditability	Full
Every step a ZenML Pro agent takes shows in its chat as it happens: messages, reasoning, the commands it runs and the tools it calls, and what it produced is recorded as artifacts such as pull requests, pipeline runs, model versions, charts and reports. Kitaru records each agent run as a session holding every model call with the requested and resolved model, tokens and cost, plus every tool call and decision. ZenML versions every pipeline step, dataset and model with lineage, and Enterprise adds audit logs. Sourcedocs.zenml.io/kitaru/guides/llm-callsread 2026-11-09
Memory & State Persistence	Partial
Each ZenML Pro chat keeps its own sandbox, and a paused chat starts again with the repository checkouts and working files it left behind, while environment snapshots let new chats skip repeated setup. In Kitaru, investigations hold reviewers' judgments pinned to exact traces, and later evaluators and replays build on them. Nothing gives an agent a memory store with a stated scope and lifetime that it writes and reads across chats; what persists is the sandbox's own state and the notes people keep. Sourcedocs.zenml.io/pro/bots/bots/bots-core-conceptsread 2026-11-09
Deployment & Data Residency	Full
Open source ZenML and Kitaru run wherever the customer installs them, from a laptop to Docker or Kubernetes with Helm against the customer's own Postgres, with workers running replays and evaluators inside the customer's environment. ZenML Pro comes as SaaS, as a hybrid with workspaces in the customer's infrastructure, or self hosted, and Enterprise runs on premises or air gapped with the customer's own LLM endpoint. On Starter the hosted agents read traces through revocable keys while data, artifacts and compute stay in the customer's cloud. Sourcezenml.io/pricingread 2026-11-09
Prebuilt Agents / Templates / Packs	Full
ZenML Pro comes with a team of hosted AI engineers, each with its own job: an eval engineer that reads every session next to the team's evaluations and writes the missing check, a reliability engineer that works out why a session or pipeline run failed and tests the fix, and a cost engineer that reviews agents and pipelines each week and tests a cheaper model. Kitaru ships ready importers for six tracing tools, recording adapters for six agent frameworks, and agent skills for Claude Code, Codex and Cursor. Sourcezenml.ioread 2026-11-09
Triggers & Channel Coverage	Full
Work reaches the agents without anyone at a keyboard. ZenML Pro automations start a bot chat with a fixed prompt on a cron schedule in any time zone, mentioning the ZenML Pro app in a Slack thread or on a GitHub issue or pull request starts a chat bound to that thread, and pipelines start on schedule triggers, event triggers or authenticated webhooks from outside systems. Hosted AI engineers read new traces as they arrive and run scheduled evaluations, with alerts when an evaluation starts drifting. Sourcedocs.zenml.io/pro/core-concepts/triggersread 2026-11-09
Model Flexibility & Routing	Full
Kitaru replays swap the model behind a recorded agent at the model call boundary, replacing every model or mapping one to another, and run on the providers and keys the customer picks, so a team tests a cheaper model on the same cohort before switching. Enterprise points the hosted AI engineers at the customer's own LLM endpoint, while on the Starter plan they run on models ZenML chooses. Typed model judges in Kitaru use TypeSafe's jev model, and deterministic checks use no model at all. Sourcedocs.zenml.io/kitaru/guides/replay-and-overridesread 2026-11-09
APIs / SDKs / MCP Extensibility	Full
ZenML has a Python SDK, a CLI and a REST API, with a separate ZenML Pro API covering organizations, roles, role assignments and permission checks, and Bots chats and artifacts can be created through the API as well. Kitaru offers Python and TypeScript SDKs, a CLI and an MCP server that gives a coding assistant typed operations over sessions, cohorts, evaluators and experiments, with capability modes for actions that create, change or delete state. Sourcedocs.zenml.io/kitaru/getting-started/setupread 2026-11-09
Testing, Debugging & Optimization	Full
Kitaru turns review into tests. Findings become versioned evaluators, either deterministic code checks or typed model judges, that run on every imported session; a check can be applied blind to sessions nobody read, with predictions locked before labels are revealed; and replays rerun a frozen cohort of failed sessions against a changed model, prompt or branch to compare outcomes side by side. The same experiment can run on every commit as a regression suite, and the hosted AI engineers test each fix against what runs today. Sourcezenml.io/product/kitaruread 2026-11-09
Browser / Computer-use	Partial
No ZenML product drives a browser or desktop for an agent. The closest reach is the Firecrawl connector, a third party service that fetches and scrapes web pages for a bot at run time, alongside the sandbox itself, an isolated computer where the agent runs commands and code against the team's repositories. Kitaru's replays work at the model and tool call boundary of a recorded agent rather than through any screen. Sourcedocs.zenml.io/pro/bots/botsread 2026-11-09

The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded

Pricing

Open source ZenML and Kitaru are free. ZenML Pro Starter is $39 a month with unlimited seats; Enterprise is quoted.

Flat monthly plan per covered AI agent and ML project, with unlimited seats; Enterprise priced on the agents, projects and pipeline runs a team actually runs

Free tierTrial available

Included quota

Starter covers one AI agent, one ML project, 100 pipeline runs a month and unlimited team members, with no limit on sessions read.

What is public

The free open source editions, the $39 Starter plan with its agent, project and run limits, and the Enterprise feature list.

Billing mechanics

Starter is a self serve monthly subscription after a 14 day trial; Enterprise is contracted through sales and priced on what a team runs.

Cost watchouts

Replays and typed model judges run on the customer's own model provider keys, and pipelines run on the customer's own compute, so both costs land on the customer's own bills.

Variable cost rationale

The platform price is flat with unlimited seats and no session limit, so the variable costs are the model calls replays and judges make on the customer's own keys and the compute pipelines use on the customer's own cloud.

Additional watchouts

Starter covers a single agent and a single ML project, so a second agent or project means an Enterprise conversation.

Overage / add-ons

More agents, projects or pipeline runs move to Enterprise pricing through sales; no per unit overage is listed for Starter.

Sales call required

Mixed (some tiers require a call)

Free / trial

ZenML and Kitaru are free open source under Apache 2.0. ZenML Pro Starter starts with a 14 day trial with full access and no card.

Lowest paid plan

ZenML Pro Starter at $39 a month for one AI agent, one ML project, up to 100 pipeline runs a month and unlimited seats

Commercial notes

ZenML GmbH keeps ZenML and Kitaru open source and sells ZenML Pro, which hosts both and adds hosted AI engineers for agents and ML pipelines.

Key ambiguities

The plan comparison text does not show which plan carries pipeline webhooks and event triggers, which the trigger docs list among paid features, and Enterprise has no published price.

Cancellation / refund

The trial needs no card; when it ends, adding a card keeps the workspace running with nothing changed but billing.

Support SLA / resale

Open source has community support, Starter has email support, and Enterprise has dedicated support with an SLA, a named engineer, workshops and architecture reviews.

Missing data

Enterprise prices, contract terms and the exact plan split for pipeline triggers and webhooks.

Agentic Index verified 2026-11-09

Official pricing page Docs

Related vendors

Alternatives to ZenML

The closest documented capability profiles to ZenML among agent infrastructure platforms tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.

Windmill12.0 / 14A lighter documented profile than ZenML
PostHog11.5 / 14A lighter documented profile than ZenML
Agno13.0 / 14Fuller documented coverage on Knowledge Grounding & RAG and Memory & State Persistence
Haystack12.0 / 14Fuller documented coverage on Knowledge Grounding & RAG
LiveKit11.0 / 14A lighter documented profile than ZenML
Mastra14.0 / 14Fuller documented coverage on Knowledge Grounding & RAG and Memory & State Persistence

Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded

Compare ZenML with other vendors →Compare all agent infrastructure platforms →See where ZenML ranks →AI agents for insurance →AI agents for banking and financial services →

Data confidence: medium