Self-Hosted AI Automation: n8n, Dify, Langflow and Flowise Compared

The gap between "I have an idea for an automation" and "it runs every morning without me" used to be a weekend of glue code and a server you'd forget to renew. Visual workflow builders closed that gap, and the AI-native ones closed it further — you can now wire an LLM into a pipeline the same way you'd wire a database.

Four projects dominate this space. They look similar in screenshots and are meaningfully different in practice. There's also a licensing trap here that catches people building businesses, so read the last section before you commit.

Star counts are from GitHub at the time of writing.

n8n — the general-purpose workhorse

n8n-io/n8n · ~203,000 stars · TypeScript

n8n is a general automation platform — think Zapier, self-hosted — that grew strong AI capabilities. Hundreds of integrations, a visual canvas, and the crucial escape hatch: any node can be arbitrary JavaScript when the visual approach runs out.

That escape hatch is why n8n tends to survive contact with real problems. Visual builders are pleasant until you need something the nodes don't cover, and then you're stuck. Being able to drop into code for one step and stay visual for the rest is the right design.

For AI specifically: LLM nodes, vector store integrations, agent nodes with tool calling. But its real strength is being good at everything around the AI — the triggers, the API calls, the error handling, the scheduling.

Licence warning: n8n is "fair-code," not open source. Self-host it freely for internal use; offering it as a hosted service to others is restricted. Read the terms if that's your plan.

Dify — built for LLM applications specifically

langgenius/dify · ~154,000 stars · TypeScript

Where n8n is automation that does AI, Dify is an LLM application platform. It assumes you're building something whose core is a model: prompt management with versioning, RAG pipelines, agents with tools, evaluation, and observability into what your prompts actually cost and return.

The evaluation and observability parts are what set it apart. Anyone who has shipped an LLM feature knows the hard problem isn't building it — it's knowing whether a prompt change made things better or worse. Dify takes that seriously in a way general automation tools don't.

If the thing you're building is an AI product, start here rather than n8n.

Licence: modified open source with restrictions on multi-tenant hosting and branding removal. Fine for internal and most commercial use; check before reselling.

Langflow — the clearest way to learn

langflow-ai/langflow · ~154,000 stars · Python · MIT

Langflow is a visual builder for LangChain-style pipelines. Drag components, connect them, watch data flow through.

Two things recommend it. First, it's MIT licensed — genuinely permissive, no asterisks, the only one of these four you can build a commercial product on without reading legal terms carefully.

Second, it's the best of these for understanding what a RAG pipeline or agent loop actually does. Seeing the chunking, embedding, retrieval and generation steps as connected boxes teaches the architecture faster than reading documentation. Even if you end up writing code, an afternoon here is a good investment.

It exports to Python, so a prototype can become a real application.

Flowise — the lightweight option

FlowiseAI/Flowise · ~55,000 stars · TypeScript

Flowise is the leanest of the four: drag-and-drop LLM chains and agents, quick to stand up, low resource footprint. Good for prototyping and for embedding a chatbot somewhere without much ceremony.

Smaller community than the others and a slower release cadence at the time of writing. Good for prototypes; think harder before making it load-bearing.

What people actually build

Concrete patterns that work, rather than the demo-ware:

Scheduled research digests. Pull from RSS, APIs or scraped pages on a schedule, summarise with a model, deliver to email or Telegram. Genuinely useful and easy to build.

Document triage. Watch a folder or inbox, extract structured fields from PDFs, write rows to a database, flag exceptions for a human. This replaces real manual work.

Support triage. Classify incoming messages, draft replies for the common ones, route the rest with context attached. Keep a human on send.

Data pipeline monitoring. Check that a job ran, that numbers look sane, that an API is responding — alert with an LLM-written explanation of what looks wrong. Boring and high value.

Content repurposing. One long piece into summaries, social posts, newsletter sections. Works well because the source material is yours.

The mistakes worth avoiding

Putting the LLM where a rule belongs. If a decision is if amount > 1000, write that. Models are for judgement, not arithmetic. Every unnecessary model call adds cost, latency and a chance of being wrong.

No human in the loop on anything outbound. Automations that send email, post publicly, or move money need review. The failure mode isn't that it breaks — it's that it works confidently and wrongly, a hundred times, before anyone notices.

Ignoring cost until the bill. A workflow firing on every webhook, each making three model calls, gets expensive quietly. Set spend limits at the provider. Log token usage from the start.

No error handling. External APIs fail. Rate limits happen. A workflow without retry and alerting is one that silently stopped working three weeks ago.

The licensing thing, properly

This catches people, so it's worth being precise. Of these four, only Langflow is MIT. The others use custom or restricted licences:

• n8n — "fair-code," restrictions on offering it as a service to third parties.
• Dify — open source with added conditions around multi-tenancy and branding.
• Flowise — non-standard licence, check the current terms.

None of this affects you running one internally, which is what most people do. It matters if you're building a product on top and expecting standard open-source freedoms. Read the LICENSE file — not the README, not a blog post — before you build a business on any of them.

Where to start

If you want general automation with AI as one capability: n8n. If you're building an AI product: Dify. If you want to understand how these pipelines work, or need a clean permissive licence: Langflow.

All four self-host with Docker in a few minutes. Build one small real thing — a daily digest of something you already check manually — rather than an elaborate workflow you'll abandon. The small one that runs every day teaches you more than the ambitious one that never quite ships.