The best open source workflow automation tools in 2026

Open source has a precise meaning: a license the Open Source Initiative has approved, unmodified. Most lists of open source workflow automation tools ignore it, which is how the most popular project on every one of them ends up mislabelled. n8n forbids offering n8n as a service to other people. Dify and Airbyte carry similar terms. All three are source-available. Seven projects below pass the test and three fail it. We read every license in its repository on 15 August 2026.

Open source workflow automation tools compared

The three tools at the bottom hold more GitHub stars than the seven above them combined. Popularity and licensing are separate questions, and this list keeps them separate.

What counts as open source

An OSI-approved license grants four freedoms: run the software for any purpose, read it, change it, and redistribute it, changed or not. MIT, Apache 2.0 and AGPL 3.0 all qualify. A license that adds a condition on top of one of them, however reasonable the condition sounds, no longer qualifies.

The condition vendors add is almost always the same one: you may not sell this as a hosted service. That protects the vendor's revenue, and for internal use it costs you nothing. It matters when you host automations for your own customers, embed the tool in a product you sell, or need the right to fork a project whose owner changes direction. Check the column before the feature comparison. It rules out more of any shortlist.

The seven that qualify

Activepieces

The closest open source match for what people want when they ask for n8n. A visual builder, several hundred connectors, and roughly 400 MCP servers for agent use. Pieces are typed TypeScript packages, so writing a connector resembles publishing an npm module. The core is MIT Expat. Two directories sit outside that grant: packages/ee/ and packages/server/api/src/app/ee.

Kestra

Apache 2.0, written in Java, and the strongest choice when workflows belong in version control. You declare them in YAML rather than drawing them, and the editor renders the YAML back as a graph. Scheduling, backfills and event triggers are first class, which suits data and platform teams more than marketing automation.

Node-RED

Apache 2.0, out of IBM's IoT work, and older than most of this category. It stays the right tool for event-driven and device-driven flows: MQTT, serial, hardware, home automation, edge gateways. Ask it to serve as a SaaS integration platform and it will disappoint you.

Windmill

A Rust backend that turns scripts in Python, TypeScript, Go, Bash or SQL into workflows, webhooks and internal UIs. The developer's answer on this list: you write the step, Windmill handles scheduling, retries, permissions and the form around it. The backend and frontend are AGPL 3.0, the language clients and the OpenFlow spec are Apache 2.0, and code behind the enterprise compile flag is proprietary. A binary built without that flag is AGPL 3.0 throughout.

Huginn

MIT, Ruby, 50k stars, and no company behind it. Agents watch sources and act: scrape a page, poll a feed, send an alert, chain the result into another agent. It predates the current category and shows its age in the interface, but the license is clean, the project is still maintained, and nobody can change the terms on you.

Temporal

MIT, and not a builder. You write workflows in Go, Java, TypeScript, Python or .NET, and Temporal makes them survive crashes, restarts and deploys by replaying their history. Reach for it when correctness over long-running processes matters more than letting non-developers edit the steps. Our durable execution post compares it with the other engines in that class.

Automatisch

AGPL 3.0, except files with .ee. in the name. It aims squarely at Zapier's job with a small connector catalog and a simple self-hosted install. Pick it when you want a short list of integrations under a copyleft license rather than a large one under a restricted license.

Popular, and not open source

Each of these three ships its source publicly, and each adds a condition that takes it out of the definition.

n8n uses the Sustainable Use License, which the project calls fair-code. Internal use is free. Offering n8n as a hosted service to third parties is not. Files with .ee. in the name or .ee in the directory name need a paid n8n Enterprise License. The license also states that branches other than main carry no grant at all, so the code you read on a feature branch is not licensed to you.

Dify uses a modified Apache 2.0. You may not run a multi-tenant environment without written permission, and you may not remove or modify the logo and copyright notices in the console.

Airbyte uses the Elastic License 2.0. You may not provide the software to third parties as a hosted or managed service, and you may not circumvent its license key.

None of this makes them bad software. n8n has the widest integration catalog of any self-hostable tool by a distance, and for internal automation the license never comes up. Call them source-available and the decision stays honest.

Open core cuts the other way too

A green license column can still hide a paywall. Activepieces, Windmill and Automatisch each keep an enterprise directory outside the open grant, the open-core pattern: the repository is open, and the feature you saw in the demo may sit in the part you cannot use. Search the repository for ee or enterprise before you plan around a feature. Kestra, Node-RED, Huginn and Temporal carry no such carve-out in their license files.

Why we chose MIT

Dafthunk is ours, and it is younger and smaller than every project ranked above, so it stays out of the list. The license choice is worth explaining, because we made it deliberately and it costs us something. Dafthunk is MIT throughout, with no enterprise directory and no clause reserving hosting to us.

We build in the open because automation software asks for an unusual amount of trust. It holds your API keys, reads your data and acts on your behalf while nobody is watching. You cannot audit a promise. You can audit a repository.

We trust users with the whole grant for the same reason. A license that says you may run this but not sell it is a bet that some of your users are competitors first. Writing that clause means writing it against everyone, including the person who wants to host one workflow for a client.

The third reason has arrived recently. Code is becoming config. When an agent can read a codebase and change it in an afternoon, the cost of adapting a tool collapses, and the binding constraint moves from skill to permission. The exe.dev blog put the case for developer tools directly in Devtools must be open source: agents make personalizing a tool routine, and a closed tool cannot be personalized at all. A workflow engine is a developer tool by that standard. If you want a node that does not exist, the fastest path should be writing it, not filing a feature request.

What this costs: anyone can take the source, host it and sell it, including against us. We accept that. If the hosted product only survives because the license forbids the alternative, it was not worth running.

Dafthunk puts a visual canvas over Cloudflare Workflows, so nodes execute as durable steps and an idle deployment costs nothing. It ties you to Cloudflare, which rules it out for air-gapped work. Browse the node reference, the workflow templates, or core concepts.

Questions people ask

Is n8n open source?

No. n8n ships under the Sustainable Use License, which the project describes as fair-code. The source is public and free for internal use, and offering n8n as a service to other people requires a commercial license. The OSI has not approved it.

Which one should I pick?

Activepieces if you want a visual builder and a permissive license. Windmill if your team writes scripts. Kestra if workflows belong in version control. Node-RED for devices and events. Temporal if durability matters more than a canvas.

Does AGPL 3.0 count as open source?

Yes. The OSI approves it. It is copyleft rather than permissive, so distributing a modified version, including running it as a network service, obliges you to publish your changes under the same license. That duty is a condition of the grant, not a restriction on who may use it.

Notes on this list

We read each project's LICENSE file in its repository and pulled star counts from the GitHub API on 15 August 2026. Licenses change, so check the file before you commit to one. We quote no prices and no connector counts, because both go stale faster than we update this page. If a claim here is wrong or has aged badly, open an issue and we will correct it.

Related reading: the best low-code workflow automation tools, which covers the hosted platforms this list leaves out, and Dafthunk compared with n8n, Zapier and Make.