Open source AI automation has stopped being a hobbyist corner. In 2026, you can self-host a workflow platform that competes with Zapier on integrations, runs agentic loops with tool calling, and costs nothing beyond your own infrastructure. The hard part isn't finding options — it's separating the projects that actually work in production from the ones that look great in a demo video and fall apart the first time a webhook retries.
This roundup ranks eight tools on three things: how much you can build before hitting a wall, how painful self-hosting actually is, and whether the agent runtime survives contact with real APIs. No GitHub-star worship, no licensing apologetics. License notes where they matter — fair-code and SSPL are not OSI-approved, and that distinction matters for some teams.
1. n8n — Score: 9.0
n8n is still the one to beat. It has 400+ integrations, a visual builder that actually scales past trivial flows, and AI agent nodes that handle tool calling without making you write a custom executor. The self-hosted Community edition is unlimited — unlimited workflows, unlimited executions, full source. The catch is the fair-code license (sustainable-use), which is not OSI-approved open source and restricts hosting it as a competing SaaS. For internal use this is irrelevant; for SaaS resellers it isn't.
Best for: Technical teams who want one platform for both integration plumbing and agentic AI workflows.
Pricing: Self-hosted free. Cloud Starter $24/mo, Pro $60/mo. Enterprise quoted.
2. Langflow — Score: 8.2
Langflow is the cleanest visual builder for LangChain-style agent and RAG pipelines. Drag-and-drop graph, live execution view, and the output is real Python you can export and run anywhere. DataStax owns it now, which has accelerated polish without (yet) compromising the open source MIT core. It's narrower than n8n — you're building AI flows, not general automation — but inside that lane it's the smoothest authoring experience.
Best for: Teams prototyping RAG and multi-agent systems who want visual iteration plus exportable code.
Pricing: Open source MIT, free self-hosted. DataStax-managed cloud available.
3. Flowise — Score: 7.8
Flowise occupies the same neighborhood as Langflow but with broader LLM provider coverage and a simpler deployment story. One Docker command and you have a working agent builder with chat memory, tool nodes, and embedding pipelines. Documentation is thinner than Langflow's and the UI has more rough edges, but the integration breadth is genuinely impressive and the project ships fast.
Best for: Solo builders and small teams who want LangChain power without writing LangChain.
Pricing: Apache 2.0, free self-hosted. Cloud plans from $35/mo.
4. Dify — Score: 7.8
Dify is the most opinionated tool in this list and it benefits from that opinion. It bundles agent orchestration, RAG, prompt management, and observability into one application — you spend less time wiring components and more time tuning behavior. The workflow editor handles branching and loops cleanly, and the built-in eval and trace tooling is better than most of its competitors ship in year three. License is a modified Apache 2.0 with a multi-tenant-SaaS restriction — fine for internal use, an obstacle for resale.
Best for: Teams shipping production LLM apps who don't want to assemble five separate tools.
Pricing: Self-hosted free. Cloud Sandbox free, Professional $59/mo, Team $159/mo.
5. Sim — Score: 7.8
Sim is the newcomer that earned its rating quickly. An open source AI workspace with 1,000+ integrations and a workflow model designed for agents from day one — not retrofitted onto an automation tool. Execution traces are first-class, the agent loop is easier to reason about than most competitors, and self-hosting is a single Docker Compose file. Still maturing — expect some breaking changes between minor versions.
Best for: Builders who want a clean agent-first platform without legacy automation baggage.
Pricing: Apache 2.0, free self-hosted. Cloud plans from $20/mo.
6. Activepieces — Score: 7.2
Activepieces is the closest open source clone of Zapier in spirit — but with AI nodes built in and a self-hosting story that doesn't punish you. The piece library is smaller than n8n's, but adding new integrations is genuinely simple because the SDK is well-designed. License is MIT for the core, which makes it the cleanest open source choice for teams that need actual OSI compliance.
Best for: Teams who want Zapier ergonomics with full ownership and a real open source license.
Pricing: Self-hosted free. Cloud plans from $25/mo.
7. Browser Use — Score: 7.2
Browser Use is the specialist. It turns an LLM into a browser-driving agent that can fill forms, click through flows, scrape sites that defeat headless scrapers, and handle login walls. Not a general automation platform — it's a library and runtime for one job, and it does that job better than the browser-control hacks bolted onto general-purpose agent frameworks. Pair it with n8n or Sim for orchestration.
Best for: Workflows that require navigating real websites with authentication, JavaScript, and dynamic state.
Pricing: MIT, free. Cloud API in early access.
8. OpenHands — Score: 7.2
OpenHands is positioned as a coding agent rather than a general automation tool, but its agent runtime is general-purpose enough to count. It runs autonomous loops in sandboxed environments, handles long-running tasks better than most alternatives, and the security model around tool execution is one of the more thoughtful in the space. Worth knowing about even if your primary use isn't code.
Best for: Engineering automation, repository maintenance, and long-running autonomous tasks.
Pricing: MIT, free self-hosted.
Comparison Table
| Tool | Score | License | Strength | Self-host difficulty |
|---|---|---|---|---|
| n8n | 9.0 | Fair-code | General automation + AI | Easy |
| Langflow | 8.2 | MIT | Visual RAG/agent builder | Easy |
| Flowise | 7.8 | Apache 2.0 | LangChain without code | Easy |
| Dify | 7.8 | Modified Apache 2.0 | Opinionated LLM app stack | Moderate |
| Sim | 7.8 | Apache 2.0 | Agent-first workspace | Easy |
| Activepieces | 7.2 | MIT | Zapier-style with AI | Easy |
| Browser Use | 7.2 | MIT | Real browser automation | Easy |
| OpenHands | 7.2 | MIT | Autonomous coding loops | Moderate |
Final Picks
If you can only run one: n8n. Its range is unmatched and the AI nodes have caught up to dedicated agent platforms. Caveat: if you need OSI-approved open source for compliance, skip to Activepieces.
If you're building LLM apps specifically: Dify for the all-in-one stack, or Langflow if you want visual prototyping that exports to real Python.
If you want pure open source with no licensing weirdness: Activepieces for general automation, Flowise or Sim for AI-specific workflows.
For specialist jobs: Browser Use when you need to drive real browsers, OpenHands when you need an autonomous coding agent that can actually finish tasks.
The honest answer for most teams is to run two: a general workflow platform (n8n or Activepieces) plus a specialist for the one job it can't do well. Trying to make one tool cover everything is how you end up with brittle workflows nobody wants to touch six months later.