Dify Review

Dify lets you build AI apps as agents, agentic workflows, and chatbots that use your own data, then publish them as web apps or call them through an API. The README lists a visual canvas for building and testing workflows, support for many proprietary and open-source models (any OpenAI-compatible API, including local models through Ollama), a prompt IDE for comparing models, a RAG pipeline from document ingestion to retrieval with text extraction from PDFs and presentations, agents with their own sandbox that can run commands, handle files, and use tools, MCP servers, or custom APIs, and LLMOps with logging, performance monitoring, and integrations for Opik, Langfuse, and Arize Phoenix. Apps can consume external MCP servers and be published as MCP servers, and there is a plugin marketplace and a CLI called difyctl for terminals, scripts, and CI.

The README states minimum hardware of 2 CPU cores and 4 GiB of RAM, plus Docker and Docker Compose v2.24.0 or later. You go to the docker folder, copy .env.example to .env, run docker compose up -d, and open the install page on localhost to initialize. A separate guide covers deploying from source. SSO and RBAC for self-hosted deployments are part of Dify Enterprise, which requires contacting sales. The documentation index at docs.dify.ai/llms.txt lists the detailed pages for app types, nodes, knowledge bases, and plugins.

The experience is a canvas on which you wire LLM calls, tools, conditions, and data sources, then test inside the same interface before publishing as a web app or an API. Non-engineers can build with the visual tools, and developers can lean on the API and plugins. This section reflects the README, license, and documentation index; we did not run Dify ourselves.

• Visual workflow canvas for LLM apps and agents
• RAG pipeline with document ingestion, text extraction, and retrieval
• Hundreds of models, including local models via Ollama and OpenAI-compatible APIs
• Agents with their own sandbox, tools, MCP servers, and custom APIs
• MCP in both directions: consume external servers and publish apps as MCP servers
• Plugin marketplace, a REST API for every feature, and the difyctl CLI
• Built-in logging and monitoring with Opik, Langfuse, and Arize Phoenix integrations

• One product covers workflows, RAG, agents, prompt tooling, and observability
• Docker Compose self-hosting with modest minimums (2 cores, 4 GiB)
• Large, active community: about 158,000 stars and commits on the day of review
• A free hosted Sandbox plan to try it without installing anything

• The license forbids using the source to operate a multi-tenant environment (a tenant is one workspace) without written authorization, which rules out reselling Dify as a multi-workspace service
• You may not remove or modify the logo or copyright information in the console or applications, which restricts white-labeling of the frontend
• Self-hosted SSO and RBAC require Dify Enterprise
• A visual builder can become hard to maintain once flows grow large
• The README documents no technical limits beyond hardware minimums, so limits live in the official docs

The Community Edition is free to self-host under the Dify Open Source License, which allows commercial use as long as you respect the multi-tenant and branding conditions, and you can obtain a commercial license otherwise. Dify Cloud is a managed service with a free Sandbox plan, and Dify Enterprise adds features such as SSO and RBAC for self-hosted deployments. The pages we reviewed list no prices for the paid tiers.

Retool and Glide are builders for internal tools and apps, and Humanloop and Vertex AI are commercial platforms for building and evaluating LLM applications. Dify sits closest to the LLM-application builders, with a visual workflow tool you can run yourself. It is not a general internal-tool or dashboard builder like Retool, so choose it when the product is an AI workflow or agent. Langflow is the nearest open-source rival, with an MIT license and a more developer-centric Python approach.

• Teams building chatbots, RAG assistants, and agent workflows with a visual tool
• Organizations that want to self-host an AI application platform for internal use
• Developers who want every app exposed as an API and optionally as an MCP server
• Teams that want built-in observability

• Companies planning to resell Dify as a multi-tenant service without a commercial license
• Anyone who must white-label the frontend by removing branding
• Teams that need self-hosted SSO and RBAC without paying for Enterprise