Windmill
Also known as: Windmill Labs, windmill.dev, Windmill AI
Open source workflow engine and developer platform: scripts in 20+ languages become flows, apps and endpoints, with AI agent steps and evals, approval steps, broad triggers and full self-hosting.
Windmill is an open source, self-hostable workflow engine and developer platform. Scripts in more than 20 languages become flows (multi-step graphs with branches, loops, retries, error handlers and approval steps), internal apps and API endpoints, and they run on a Rust job engine across worker groups. Flows start from cron schedules, webhooks and HTTP routes, email, WebSocket, Postgres change capture, Kafka, NATS, Amazon SQS, MQTT, AMQP, Google Cloud Pub/Sub and Azure triggers. Any step can suspend until approvers resume it through generated URLs or an approval page with forms.
For AI work, AI agent steps run inside flows with a provider and model the customer picks (OpenAI, Anthropic, Google, Mistral, DeepSeek, Groq, OpenRouter and others), Windmill scripts as tools, nested agents, and conversation memory. Reusable agents carry evals: datasets of cases scored by AI judges or scripts, with every run kept for comparison. An MCP server exposes a workspace's scripts and flows to Claude, Cursor and other clients, and AI sessions let a chat assistant build scripts, flows and apps as drafts. The Hub offers ready-made projects and integration scripts for hundreds of services.
The Community Edition is free to self host. Windmill Cloud's Team plan is $10 per seat a month. The Enterprise Edition, from $120 a month self hosted, adds audit logs, SAML and SCIM, OpenTelemetry tracing and a SOC 2 Type II report.
Vendor details
Canonical URL
https://www.windmill.dev
Category
Agent infrastructure
Company status
independent
Use cases & customers
Target customers
Deployment options
In practice
A Python script needs to be a scheduled job, an internal tool with a UI and a webhook API all at once. Windmill generates the interface and the endpoint from the script's parameters, so one script serves all three.
An AI workflow must stop for human sign off before it acts. Any flow step can suspend for approval, holding the flow until approvers resume it through generated URLs or an approval page with forms.
Compliance rules out a managed automation service. The platform self hosts on your own infrastructure with role based access control, secret management and audit logs. Audit logs, SAML and SCIM come with the Enterprise Edition.
Sources & related URLs
Agentic Index coverage score
12.0 / 14 capabilities · 86%
| Integrations & Tool Calling | Full |
|---|---|
|
The Hub catalogs integration scripts for hundreds of services, among them Slack, Jira, Gmail, Google Sheets, Postgres, Airtable, OpenAI and Anthropic, and resource types store the credentials those scripts use. Any Windmill script can be attached as a tool to an AI agent step, so an agent takes authenticated action in outside systems with stored credentials. AI sessions can also connect to external MCP servers as clients. Sourcehub.windmill.devread 2026-09-23 |
|
| Workflow Orchestration | Full |
|
Windmill is a workflow engine whose flows compose scripts in 20-plus languages into multi-step graphs with branches, loops, early stops, retries, error handlers and suspend steps. AI agent steps run inside those flows with tools, and nested agents can be configured as tools of another agent. Sourcewindmill.dev/docs/flows/flow_editorread 2026-09-23 |
|
| Knowledge Grounding & RAG | Not documented |
|
There is no retrieval index that Windmill maintains over the customer's documents. Its persistent storage, which covers data tables, DuckLake, object storage, states and resources, holds data the customer's own scripts read and write, and any retrieval a customer builds runs through tools or providers called from a script. Sourcewindmill.dev/docs/core_concepts/persistent_storage/states_resources_shared_directoryread 2026-09-23 |
|
| Human Oversight & Guardrails | Full |
|
Any flow step can be set to suspend, a durable approval wait built into Windmill, holding the flow until the required number of approval events arrive through generated resume and cancel URLs or the approval page, with forms for the approver to fill in and a timeout. Sourcewindmill.dev/docs/flows/flow_approvalread 2026-09-23 |
|
| Security, Identity & Governance | Full |
|
For the Enterprise Edition, Windmill is SOC 2 Type II compliant, with the report downloadable from the customer portal and a trust center at trust.windmill.dev. On Enterprise, SAML and SCIM authenticate and provision users through the customer's identity provider. SSO is available on every edition, capped at 10 users on the free edition, and developer and operator roles separate who can edit from who can only run. Audit logs record every operation with side effects. Sourcewindmill.dev/docs/advanced/saml_and_scimread 2026-09-23 |
|
| Observability & Auditability | Full |
|
The Runs page lists past and future jobs with each job's logs, inputs, results and per-step detail for flows, and AI agent steps report their tool calls and token usage. A separate audit log captures every operation with side effects and the user responsible, with a configurable retention period on self-hosted instances, and an OpenTelemetry guide exports traces, metrics and logs to the customer's own observability stack. Sourcewindmill.dev/docs/core_concepts/audit_logsread 2026-09-23 |
|
| Memory & State Persistence | Partial |
|
The AI agent step keeps conversation memory that the agent reads as context across runs. In auto mode Windmill keeps up to a set number of the most recent messages, in manual mode the caller supplies the history in OpenAI message format, and memory also works with webhook-triggered runs. Windmill gives no retention period or purge path for that memory and no screen of its own for reviewing or editing what the agent remembers. States and resources keep small JSON values between job executions for the code to read. Sourcewindmill.dev/docs/core_concepts/ai_agentsread 2026-09-23 |
|
| Deployment & Data Residency | Full |
|
Windmill is open source and self-hosts in the customer's own infrastructure, with setup guides for Docker, Kubernetes with Helm, AWS EKS and ECS, and Azure. Windmill Cloud and dedicated cloud instances for Enterprise are also offered. Sourcewindmill.dev/docs/advanced/self_hostread 2026-09-23 |
|
| Prebuilt Agents, Templates & Packs | Full |
|
The Hub is a browsable catalog of assets a customer adopts into a Windmill workspace. Its ready-made projects include a support engine that drafts replies into Gmail, a recruiting app with AI CV screening, a Sentry engine that investigates recurring errors and opens a pull request with the fix, a helpdesk and an uptime monitor. It also holds integration scripts and flows for hundreds of services, which can be picked from inside the editors, and a Private Hub serves an organization's own approved set. Sourcehub.windmill.devread 2026-09-23 |
|
| Triggers & Channel Coverage | Full |
|
Scripts and flows start on their own from cron schedules, webhooks and HTTP routes, email triggers, WebSocket and Postgres change capture, and from Kafka, NATS, Amazon SQS, MQTT, AMQP, Google Cloud Pub/Sub and Azure triggers, plus native triggers, each one firing when its event arrives. Sourcewindmill.dev/docs/triggersread 2026-09-23 |
|
| Model Flexibility & Routing | Full |
|
Each AI agent step uses a model the customer picks. Providers include OpenAI, Azure OpenAI, Azure AI Foundry, Anthropic, Mistral, DeepSeek, Google Gemini, Groq, OpenRouter, Together AI and custom endpoints, with the credentials stored as reusable resources. A reasoning-effort setting is available for models that support it, and nested agents set their own provider and model. Sourcewindmill.dev/docs/core_concepts/ai_agentsread 2026-09-23 |
|
| APIs, SDKs & MCP Extensibility | Full |
|
Outside callers can use a REST API with a published OpenAPI specification, the wmill CLI with sync to Git, and TypeScript and Python clients. Webhook and HTTP route endpoints are generated for any script or flow, and an MCP server exposes a workspace's scripts and flows as tools, along with API operations on jobs and other objects. Sourcewindmill.dev/docs/core_concepts/mcpread 2026-09-23 |
|
| Testing, Debugging & Optimization | Full |
|
An evaluation harness for agents ships as a beta. AI agent evals attach a dataset of cases, each a question with its expected answer, to a reusable agent, run the agent against every case, score the answers with AI judges or scripts, and keep every run so the next one is compared against it. The point is to show whether the agent got better after each change. Sourcewindmill.dev/docs/core_concepts/ai_agent_evalsread 2026-09-23 |
|
| Browser & Computer Use | Partial |
|
Browser automation runs in an environment Windmill hosts, where a worker group installs Chromium on start and takes jobs tagged chromium. Playwright and Puppeteer examples cover testing and scraping, and a script using them can be exposed as a tool to an AI agent step. The browser control comes from those libraries in the customer's own script, and Windmill has no API of its own for screenshots, clicks or typing. Sourcewindmill.dev/docs/advanced/browser_automationread 2026-09-23 |
|
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
Product updates
AI agent steps in Windmill flows gain tool approval and a question tool in version 1.850.0. An agent can now wait for a person to approve a tool call before it runs, and can put a question to a person in the middle of a run.
Bears on: Human approval / guardrails
View sourceA new AI decision step in Windmill flows asks a decision model, TypeSafe's Jev or Cloudflare's Clef on Workers AI, typed choice, score or yes or no questions about a state and returns calibrated answers with probabilities instead of text. The editor can insert a branch for each answer option, and the step can be handed to an AI agent as a tool.
Bears on: Agent capability
View sourceAI agents are now their own item in Windmill, listed beside scripts, flows and apps with dedicated pages for running and editing them, including drafts, deployment and version history.
Bears on: Agent capability
View sourcePricing
Community Edition is free to self host. Cloud Team is $10 per seat a month, and self hosted Enterprise starts at $120 a month.
Windmill Cloud bills per seat per month. Self hosted Enterprise bills developer seats, operator seats at half price and compute units per worker, with a standard worker at $50 a month. App guests beyond 100 a month count as seats.
What is public
The Community Edition (free to self host), Cloud Team at $10 per seat a month and self hosted Enterprise from $120 a month, with its developer, operator and worker rates, are public, as are the Pro and nonprofit terms. Cloud Enterprise is custom.
Billing mechanics
Windmill offers two routes. The open source edition is free to self host on any infrastructure, where the only cost is compute and operations, and self hosted Enterprise adds seat and worker charges. Windmill Cloud starts with a free tier for small projects and scales through paid tiers adding team collaboration, priority support and higher execution limits, with enterprise terms negotiated.
Cost watchouts
Audit logs, SAML and SCIM, the SOC 2 report, OpenTelemetry tracing and longer job retention sit in the Enterprise Edition. The free edition keeps job run details for up to 30 days. Self hosted Enterprise bills compute units per worker on top of seats.
Variable cost rationale
Seats are the main meter. Cloud Team is $10 per seat a month, and app guests beyond 100 a month count as seats. Self hosted Enterprise also bills compute units per worker, with a standard worker at $50 a month, so adding workers raises the bill alongside seats. Cloud tiers carry execution limits that are not published per tier. The seat and worker rates on Team and self hosted Enterprise are fixed and published, so the bill grows with headcount and worker count rather than with runs. The open source Community Edition meters nothing and allows unlimited executions. Its only cost is the compute and operations to run it.
Additional watchouts
Audit logs, SAML and SCIM and the SOC 2 report are in the Enterprise Edition only, so a regulated or security sensitive workload on the open source edition runs without them or moves to Enterprise pricing. Which tiers include secrets management, roles and approver identity is not set out.
Sales call required
Mixed (some tiers require a call)
Free / trial
Free open source Community Edition to self host, and a one month Enterprise trial
Lowest paid plan
Cloud Team at $10 per seat per month
Commercial notes
Windmill positions itself against managed automation platforms on control. For teams that already run infrastructure, the open source edition costs effectively nothing beyond compute and operations.
Key ambiguities
Self hosted Enterprise cost depends on the worker configuration as well as seats, and Windmill's pricing calculator simulates it. Cloud Enterprise is custom.
Support SLA / resale
Community support on Discord for the free edition. Enterprise adds 24/7 priority support with a 3 hour response, a dedicated Slack channel and an SLA.
Missing data
Cloud Enterprise pricing (custom on seats and compute units) and execution limits per tier.
Related vendors
- AgentOps — Agent observability and debugging platform: open source SDKs trace…
- AgentPhone — AgentPhone gives the AI agents a team builds their own US or…
- Agno — Python agent framework and AgentOS runtime (formerly Phidata) for…
- AIsa — Resource and payment gateway for AI agents, with one key to 110+…
- Albato — Albato is a no code integration platform that sells an embedded,…
- AlphaBitCore — AI control plane for regulated financial firms: one gateway enforces…
Alternatives to Windmill
The closest documented capability profiles to Windmill 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.
- Bernstein10.5 / 14A lighter documented profile than Windmill
- LangSmith11.5 / 14Fuller documented coverage on Memory & State Persistence
- OpenRouter10.5 / 14A lighter documented profile than Windmill
- Agno13.0 / 14Adds documented Knowledge Grounding & RAG
- Anchor Browser11.0 / 14Fuller documented coverage on Browser & Computer Use
- Braintrust10.0 / 14A lighter documented profile than Windmill
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded