Langflow
Visual Builder for AI Agents, MIT-Licensed and Yours to Run
Visual workflow builder, knowledge bases and agents for LLM apps, model-agnostic and self-hostable. Free to run yourself — with two licence conditions that decide whether it fits your business.

Dify is a platform for building applications on top of language models. Not a chatbot and not an assistant — the thing you use to make one. It gives you a visual workflow builder, a knowledge base with retrieval, agent tooling, prompt management and the operational plumbing around all of it, and you can run the whole thing on your own servers.
The gap it fills is specific. Between calling an API yourself and buying a finished product there is a large amount of work that every team rebuilds identically: chunking documents, managing retrieval, versioning prompts, chaining steps, logging what happened, letting a non-engineer adjust the wording. Dify is that layer, already built.
It is also model-agnostic, which is more valuable than it sounds in a year when products get discontinued. Your application is built against Dify, not against one vendor's endpoint.
A visual canvas where steps connect: a model call feeds a condition, which branches to a retrieval step, which feeds another model call, which triggers an action. Complex behaviour becomes something you can look at and correct, which matters enormously when the person who needs to adjust it is not the person who built it.
Upload documents and the platform handles the parts that are tedious to get right — chunking, indexing, retrieval, ranking. This is the machinery behind any assistant that answers from your own material rather than from the internet, and building it yourself is a fortnight you do not get back.
Agents that use tools to complete tasks, and triggers that start a workflow on an event rather than on a person clicking something. The trigger allowances are a real part of the pricing, which tells you how central automated execution has become to the product.
Logging, annotation, prompt iteration, monitoring what your application actually did. This is the unglamorous half and the reason teams stay: an LLM application without observability is one you cannot debug and therefore cannot maintain.
Dify is open source in the sense that matters most — the code is public and you can run it — but the licence is a modified Apache 2.0, and the modifications are precisely where a business plan might collide with it.
Read that carefully against your intentions. Running Dify inside your company to power internal tools or a product backend is free and unencumbered. Reselling it as a multi-tenant service, or shipping a console with your own branding, is the business Dify itself is in — and that is what the Enterprise tier's commercial licence authorisation buys.
This is a coherent and honest business model, not a trap. But it is the single question to answer before you commit, because discovering it after launch is expensive.
Two ways to run it: yourself, or on Dify's cloud. Self-hosting from the community edition is free of licence fees, and you pay for servers and the model APIs you call. The cloud is billed per workspace per year, which is an unusual shape and worth reading twice.
| Plan | Price | Message credits | Members and limits |
|---|---|---|---|
| Community (self-hosted) | Free | Your own model costs | Single workspace, core features from the public repository |
| Sandbox | Free | 200 total | 1 member, 5 apps, 50 documents, 50MB, 30 days of logs |
| Professional | $590 per workspace / year | 5,000 a month | 3 members, 50 apps, 500 documents, 5GB, unlimited log history |
| Team | $1,590 per workspace / year | 10,000 a month | 50 members, 200 apps, 1,000 documents, 20GB, unlimited triggers |
| Enterprise | By quote | Custom | Multiple workspaces, SSO, commercial licence authorisation, SLAs, official support |
Annual billing saves about seventeen per cent against paying monthly. The number to look at hardest is the member cap: three people on Professional. A team of four is pushed to a plan costing nearly three times as much, which is a steep step and the most common reason people end up self-hosting instead.
The Sandbox tier is a demonstration rather than a plan — two hundred credits total, not per month. It is enough to understand the product and nothing more, which is fair given that the free option people actually use is running it themselves.
The strongest case, and the one the licence is most comfortable with. Self-host it, point it at your documents and your model provider, and let several teams build assistants without each of them writing retrieval code.
Because you can run the whole platform yourself, the documents in your knowledge base stay on your infrastructure. Pair it with a model you also host and nothing leaves the building at all.
Chunking strategy, retrieval quality, prompt versioning, logs. None of it is intellectually interesting and all of it takes weeks. Starting from a platform that has already made those decisions is usually the right call.
The visual builder means a product manager or a subject expert can adjust a prompt or reorder a workflow without a deployment. For applications where the domain knowledge sits outside engineering, that shortens the loop considerably.
The community edition is free to self-host, and there is a free Sandbox tier on the cloud with two hundred credits in total. Paid cloud plans start at $590 per workspace per year.
It is a modified Apache 2.0 licence. The code is public and commercial deployment as a backend or internal platform is permitted, but you may not run a multi-tenant service or remove the console branding without written authorisation.
Not one where each customer gets an isolated workspace — that is the multi-tenant case the licence restricts. Talk to Dify about a commercial licence, which is what the Enterprise tier includes.
It is model-agnostic and wraps many providers, including models you host yourself. That is a large part of its value: your application is built against the platform rather than against one vendor.
The cloud's unit of usage, allocated monthly on the paid tiers. It is separate from what the underlying model provider charges you, so plan for two bills rather than one.
Faster to start and much faster for non-engineers to adjust. Against that, you inherit someone else's decisions about retrieval and orchestration, and a large workflow on a canvas eventually becomes harder to maintain than the equivalent code.
Dify is the most complete open platform for turning language models into applications, and self-hosting it costs nothing but infrastructure. For an organisation building internal tools or a product backend, it removes weeks of work nobody enjoys doing twice.
Answer the licence question first. Internal use and backend deployment are free and clear; multi-tenant resale and white labelling are not. And if you are looking at the cloud, check the member cap before the price — three people on the entry plan is what usually decides the matter.