Dify: Build AI Apps Visually — Workflows, RAG and Agents in One Platform

Dify: Build AI Apps Visually — Workflows, RAG and Agents in One Platform

Every team now wants an AI application. A chatbot that answers customers from the company's own documents. A workflow that reads a pile of PDFs and returns a summary. An agent that checks a calendar, searches the web and drafts a reply. The idea is easy to describe and surprisingly hard to build, because the finished product is never just a model: it is a model plus a document pipeline plus a user interface plus a way to publish it plus a way to watch it work.

Dify is a platform that bundles all of that into one place, and lets you build it by dragging blocks on a canvas instead of writing the plumbing yourself. It is maintained by LangGenius, it was founded in March 2023 by Luyu Zhang, and it has grown into one of the most-starred AI platforms in the world — about 161,000 stars as of 16 November 2026. Its newest release, version 1.20.0, arrived on 18 November 2026. The name is short for "Do It For You", and that is an honest description of what it tries to do.

But one fact must be understood before anything else, because most introductions get it wrong. Dify is not plain open source. It is released under the Dify Open Source License, a modified version of the Apache License 2.0 with extra conditions attached. The source code is public to read and run, and commercial use is allowed — but you may not run it as a multi-tenant service for your own customers without written permission, and you may not remove the Dify logo and copyright information from the interface. This post states that licence precisely, because the difference between "open source" and "source-available with conditions" is the difference between a free tool and a legal problem.



What Is Dify, in Plain Words?

Dify is an application development platform for large language models. Here is the simple version: it is a workshop where you build AI applications by connecting blocks on a screen, and it supplies everything the application needs to run — the model connection, the document search, the chat interface, the publishing step and the logs.

You start in the Studio, where you choose what kind of application to build. A Chatbot is a conversational application with memory: a customer asks a question, the application answers, and it remembers the conversation. A Workflow is a one-off task: data goes in, a result comes out, useful for batch processing, classification and report generation. A Chatflow is a workflow that runs on every message in a conversation, so it can keep variables and memory between turns. An Agent is an application where the model decides for itself which tools to call. A Text Generator is a simple application that turns an input into a piece of text. Five types, one canvas, and the same building blocks underneath all of them.

The reason Dify has grown so fast is that it covers more of the finished product than its rivals. A code-first library gives you the pieces and leaves the assembly to you. Dify gives you the whole assembly line: model management for more than 100 providers, a knowledge base that turns your documents into searchable material, a visual workflow builder, an agent framework, a plugin marketplace, and a way to publish the finished application as a web page, an embedded widget, or an application programming interface. You can build a working support chatbot in an afternoon, without writing a backend.

The Licence Warning: Modified Apache 2.0, Not Plain Open Source

Dify is released under the Dify Open Source License, which its own licence file describes as a modified version of the Apache License 2.0 with additional conditions. Everything in the Apache License 2.0 applies except where those conditions change it, and the conditions are the part that matters. Commercial use is allowed, including using Dify as a backend service for other applications or as an application development platform for enterprises. That covers most readers of this blog: building an internal chatbot for your own company, or an application that your own team uses, is fine.

The first condition is the restriction on multi-tenant service. Unless Dify authorises you in writing, you may not use the source code to operate a multi-tenant environment — in practice, one tenant means one workspace, so running separate workspaces for separate customers as a hosted service needs a commercial licence from LangGenius. The Enterprise tier's feature list includes commercial licence authorisation, and that is where this permission lives. If your plan is to build a business by selling access to Dify itself, stop and read the licence before you write a line of code.

The second condition concerns branding: while using the Dify frontend, you may not remove or modify the logo or copyright information in the console or in the applications. The paid Premium edition, sold through the AWS and Azure marketplaces, is the route that allows removing the "Powered by" branding. The third condition concerns contributors: by contributing, you agree that the producer can adjust the open-source agreement to be more strict or more relaxed as it sees necessary, and that your contributed code may be used for commercial purposes, including in the cloud business.

Put plainly, three uses are fine and one is not. Running Dify on your own server for your own company's work is fine. Running it at home for personal projects is fine. Using it as the backend of your own application is fine. But offering Dify itself to your customers as a hosted, multi-tenant service — where each customer gets their own workspace and you charge them for it — needs written permission and a commercial licence. That is the line, and it is the reason this post never calls Dify plain "open source" without the qualification.

How the Visual Canvas Actually Works

Open a Workflow in Dify and you see a blank canvas with a menu of nodes. You drag a Start node onto the canvas, then drag the next node beside it, and a line connects them. Each node does one job, and data flows along the lines from one node to the next until the workflow reaches an End node. It is the same idea as the workflow canvas in the n8n post published on Tech Know on 17 November 2026, but the nodes here are built for language models rather than for office automation.

The node types are the vocabulary of the platform. An LLM node sends a request to a model. A Knowledge Retrieval node searches your knowledge bases for the documents relevant to a question. A Question Classifier routes a request by topic. An If-Else node branches the flow. An Iteration node repeats a step over a list. A Code node runs Python or JavaScript in a sandbox. A Template node reshapes text with Jinja2 templates. An HTTP Request node calls an external service. A Parameter Extractor pulls structured fields out of free text. A Document Extractor reads files. An Agent node drops an autonomous agent into the middle of a workflow. A Human Input node pauses the flow and waits for a person to answer.

The Knowledge Base: Turning Your Documents Into Answers

The knowledge base is where Dify earns its keep, and it is the feature that separates it from a simple chatbot builder. It is the retrieval part of RAG — retrieval-augmented generation — the technique where a model answers from your documents instead of from what it memorised during training. Upload your documents, and Dify handles the whole pipeline: it extracts the text, splits it into chunks, converts the chunks into vectors with an embedding model, stores them, and searches them when a question arrives.

Building one takes minutes. Go to the Knowledge tab, create a knowledge base, and upload documents — PDF, Word, PowerPoint, Markdown, plain text, web pages, even Notion pages, up to 50 MB per file. Configure the chunking and the indexing method, connect an embedding model, and the documents become searchable. The Knowledge Retrieval node in a workflow then takes a question, pulls the relevant chunks, and hands them to the model as context. The model answers from your documents, not from its training data, which is what stops it inventing facts about your company.

Retrieval uses hybrid search, combining vector similarity with keyword matching, which performs better than pure vector search for technical content where exact terms matter. You can choose your vector store: Weaviate is bundled with the self-hosted stack, and Qdrant, Milvus, Pinecone, pgvector and Chroma are supported for larger deployments. Knowledge bases are also queryable through the API, so your existing applications can use Dify as an external retrieval service without going through the full workflow builder. One honest warning belongs here: a knowledge base is only as good as the documents inside it. Out-of-date price lists and contradictory policy documents will produce confident, wrong answers — and the model will not know which document to trust.



Agents: When the Application Starts Deciding for Itself

Beyond linear workflows, Dify supports agents — applications where the model decides for itself which tools to call, in what order, until the job is done or it runs out of steps. The Agent node gives a language model planning capability, using reasoning strategies such as chain-of-thought, and the agent can call more than 50 built-in tools: web search, a code interpreter, image generation, HTTP requests, Wikipedia, weather, and more.

Version 1.18.0, released on 23 October 2026, added Agent Output Routes, where an agent node chooses which workflow branch runs next instead of the branch being fixed in advance. The same release added portable .ifpkg app packages — a whole application, with its resources, packed into one file you can move between installations — and a beta Agent Knowledge Base, plus a new command-line client called difyctl that creates, imports, exports and publishes Workflow and Chatflow apps from the terminal.

Version 1.19.0, released on 4 November 2026, added nested workflow tracing, so the steps inside a Workflow Tool now appear nested under the invoking tool in observability platforms such as Langfuse and LangSmith. It also exports transitive workflow dependencies in package format version 2, and bundles plugin daemon 0.6.14-local.

Version 1.20.0, released on 18 November 2026, exports traces to any OpenTelemetry collector and keeps Langfuse working.

The newest way to build an agent is the Agent page, marked Beta, which arrived with the 1.19 line. You give the agent a name, an optional role and a description, and it opens in a Configure panel: model, prompt, skills, reference files, tools, knowledge retrieval and environment variables. The agent runs in a Linux sandbox, sized at 20 GB of storage, where it can install programs and handle files, so it can take on open-ended work rather than only the tools you configured. Skills are uploaded SKILL.md packages that teach the agent a repeatable procedure, and a Build mode lets you describe the agent in a chat while it fills in the prompt, skills and files itself.

Models and MCP: More Than 100 Providers, Two-Way Tooling

Dify has no model of its own, and it does not lock you into one provider. Models are connected through a plugin marketplace: after first login, go to Settings, then Model Provider, install the provider plugin and add your key. OpenAI, Anthropic, Google Gemini, DeepSeek, Mistral, Cohere, xAI, Tongyi and dozens more are available, along with Azure OpenAI and AWS Bedrock, and self-hosted models through Ollama or any OpenAI-compatible inference server. The Ollama post published on Tech Know on 13 November 2026 explains the local-models idea in detail: run a model on your own computer, connect it to Dify, and your chatbot costs only electricity.

In July 2026 Dify added native bidirectional MCP support — the Model Context Protocol. That means Dify agents can both call external MCP servers and expose their own tools as MCP endpoints, a two-way door that turns Dify into a hub in the fast-growing MCP tool ecosystem. An agent inside Dify can use tools from any server that speaks MCP, and tools built inside Dify can be called by agents running elsewhere. For teams building on the MCP standard, that is a genuine reason to choose Dify over a platform that only consumes tools.

What Does Dify Cost? The Full Price Picture

Dify ships two ways: a managed Cloud service and a self-hosted edition you run on your own infrastructure. The Cloud plans are priced per workspace, not per person, and the subscription governs the whole team's member limit, feature access and usage quotas. The self-hosted Community Edition is free under the Dify Open Source License, with all core features released under the public repository, on a single workspace. What the paid tiers sell, therefore, is narrower than usual: managed hosting, higher limits, more team members, and — at Enterprise — multiple workspaces, single sign-on and commercial licence authorisation.

PlanPriceMessage creditsMembersAppsKnowledge documents
Sandbox (Cloud)Free200, one time only1550
Professional (Cloud)$59 per workspace per month, or $590 per year5,000 per month350500
Team (Cloud)$159 per workspace per month, or $1,590 per year10,000 per month502001,000
Community (self-hosted)FreeYour own model keysUnlimited by licenceUnlimited by licenceUnlimited by licence
EnterpriseCustom quoteCustomCustomCustomCustom

Read the Sandbox allowance carefully: 200 message credits, with no "per month" attached, where both paid tiers explicitly say per month. It is a trial allowance, not a monthly one, and it runs out. Annual billing saves about 17 per cent — Professional is $590 a year against $708 paid monthly, Team $1,590 against $1,908. Prices exclude tax, and the pricing page was checked on 4 November and 9 November 2026 by independent guides; re-check it at publish time, because cloud prices change without announcement. Students and teachers with a school email can get the yearly Professional plan free for a year.

The honest catch is the message credits. The cloud credits buy very little model use: at list prices, Professional's 5,000 monthly credits are worth only a few dollars of the same answers on your own API key, so the $59 mostly pays for seats, apps, storage and hosting, and real volume runs on your own key anyway. That is not a flaw — it is the business model, and it means the Cloud fee is really a hosting fee. If your team already runs Docker in production, the self-hosted Community Edition costs you only the server.

Self-Hosting: Docker Compose and the Real Requirements

Self-hosting Dify is deliberately simple to start and heavier than the readme suggests once it is running. The standard route is Docker Compose: clone the repository, copy the environment file, edit your credentials, and start the stack. The documented minimum is 2 CPU cores and 4 GiB of RAM, plus Docker Compose 2.35.0 or later — the 1.19.0 release raised that requirement — and OpenSSL. Replace the default credentials before anyone else can reach the installation; the default passwords are public knowledge.

A hands-on run of version 1.19.0 on 12 November 2026, reported by an independent reviewer, found the reality behind the minimum: about 15 long-running containers and 11.4 GB of images, no model provider installed until you add one from the marketplace, and anonymous telemetry switched on by default. For source installs rather than Docker, the 1.18.0 release notes require Python 3.13 for the API, Node.js 24.22.1 or later in the 24.x line, and pnpm 12.10.1 — and after upgrading, the difyctl client must be reinstalled and you must sign in again.

Upgrading needs care. The 1.18.0 upgrade guide warns to upgrade the API, workers, Agent Backend and Agent runtime images together, to deploy the matching 1.18.0 Agent Backend and Local sandbox images, and to preserve your PASSWORD_PEPPER configuration — losing it locks you out of existing data. The 1.19.0 release adds a further rule: rolling upgrades must route tracing to the new ops_trace_v2 queue while old ops_trace workers drain. None of this is difficult, but all of it is the kind of operational work that a Cloud subscription exists to remove. Choose self-hosting when someone on your team already runs Docker in production; otherwise the cloud fee is mostly buying that person's time.

Dify Against Langflow, RAGFlow and n8n: An Honest Comparison

Most readers will compare Dify with Langflow, RAGFlow and n8n, the three closest alternatives. All four can build AI applications, and the differences are about scope, licence and who does the operating.

DifyLangflowRAGFlown8n
What it isFull LLM app platform: workflows, RAG, agents, publishingVisual flow builder for LLM appsRAG-focused document platformWorkflow automation with AI agent nodes
GitHub stars (Nov 2026)About 161,000About 157,000About 94,300About 210,500
LicenceDify Open Source License (modified Apache 2.0, with conditions)MITApache 2.0Sustainable Use License v1.0 (fair-code)
Multi-tenant SaaS from the codeNeeds written permissionAllowedAllowedNot allowed without a commercial agreement
Built-in RAG knowledge baseYes, with hybrid searchBasicYes, its core strengthNo
Self-host minimum2 vCPU, 4 GB RAM1 vCPU, 1.5 GB RAMHeavierLight, one container
Free hosted tierYes, SandboxNoNoNo, trial only

Who Should Use Dify — and Who Should Not

Dify suits teams who want to ship a real AI application — a support chatbot, a document assistant, a multi-step workflow, an agent with tools — without assembling the plumbing themselves, and who either have someone comfortable with Docker or are willing to pay the Cloud fee to avoid it. It suits a company with documents worth searching, a team that wants to see every step a workflow took, and a product that needs to be published as a web page, a widget or an API rather than as a script on someone's laptop.

It does not suit three kinds of user. The first is anyone planning to resell Dify itself as a multi-tenant hosted service — the licence forbids it without written permission, and Langflow or RAGFlow would be the honest choice instead. The second is anyone who wants a single-purpose tool: if all you need is a chatbot over one PDF, a lighter tool will get you there faster than a fifteen-container platform. The third is anyone with no one to operate it: a self-hosted Dify that nobody updates is a security risk, not an asset, and in that case the Cloud Sandbox or Professional plan is the safer way to start. The ComfyUI post published on Tech Know on 16 November 2026 makes the same point about a different kind of platform: the tool is only as good as the person keeping it running.

Frequently Asked Questions

Is Dify open source? Not in the plain sense. Dify is released under the Dify Open Source License, a modified Apache 2.0 with extra conditions. The code is public and commercial use is allowed, but running a multi-tenant service needs written permission and a commercial licence, and the Dify logo and copyright in the frontend may not be removed. Call it source-available under a modified Apache 2.0, not plain open source.

What is the latest version of Dify? Version 1.20.0, released on 18 November 2026, is the newest release. It adds OpenTelemetry trace export and a Langfuse compatibility fix, and it requires Docker Compose 2.35.0 or later.

How many stars does Dify have? About 161,000 as of 16 November 2026, based on live repository data. It passed 100,000 stars in October 2025 and 160,000 in late October 2026.

Is Dify free? The self-hosted Community Edition is free, with all core features, on a single workspace, under the Dify Open Source License. Dify Cloud has a free Sandbox plan with 200 one-time message credits. Paid Cloud plans start at $59 per workspace per month, or $590 a year, for Professional.

Can I use Dify with Ollama and local models? Yes. Dify connects to Ollama and to any OpenAI-compatible inference server, so a local model can power a workflow, a chatbot or an agent. No model is bundled with a fresh install — you install a provider plugin from the marketplace first.

The Bottom Line

Dify is the most complete visual platform for building AI applications that you can run yourself — workflows, knowledge bases, agents, model management, publishing and observability in one console, with about 161,000 stars and a release every month to prove that people actually use it. Version 1.20.0, from 18 November 2026, makes it easier to trace what agents do and easier to move applications between installations. For a team with documents worth searching and someone willing to operate the stack, the self-hosted Community Edition costs nothing but the server. For a team without that person, the Cloud plans start at $59 a month and buy the operations back.

By Umar Farooq, Tech Know.

Post a Comment

0 Comments