Tool
Dify
An open-source platform for building LLM apps — visual workflows, RAG pipelines, agents, and model management — that you can self-host as Community Edition or try on Dify Cloud.
What it is
Dify (repo langgenius/dify) combines a console UI with orchestration pieces: AI workflows, RAG, agents, model providers, and observability hooks (integrations called out upstream include Opik, Langfuse, and Arize Phoenix). Community Edition is meant to be Docker-compose self-hosted; LangGenius also runs a hosted Cloud product and sells enterprise features. You still bring model API keys or point at local backends — the platform is glue and product surface, not free frontier inference.
What it replaces
It replaces stitching raw LangChain scripts and a homegrown admin UI when you want a visual path from prototype chat/RAG/agent to something you can hand to a small team. Self-host when data residency or cost control matters; Cloud when setup time matters more.
It does not replace careful eval of retrieval quality, prompt failure modes, or model bills. It is not “plain Apache, rebrand and sell as multi-tenant SaaS” — see license below. Frontend branding rules apply if you ship the console UI.
License
Dify uses a modified Apache License 2.0 (“Dify Open Source License”), verified from the repo LICENSE. Extra conditions (paraphrased — read the file for legal text):
- Multi-tenant service: operating a multi-tenant environment (one tenant = one workspace in Dify’s definition) needs written authorization / commercial license from the producer.
- Logo / copyright: you may not remove or modify logo or copyright notices in the Dify frontend (console/apps). API-only use without the frontend is carved out of that branding rule.
- Contributors agree the producer may adjust the open-source agreement and may use contributions commercially (including cloud).
Internal commercial use as a backend or enterprise app platform is contemplated; SaaS-shaped multi-tenant clones are the restricted path. Re-check the LICENSE before compliance decisions — producers can tighten or relax terms.
Where it falls over
- Self-host is a real stack (API, workers, DB, vector store, web) — Docker compose is “quick start,” not zero ops.
- Model quality and cost still come from whatever providers you wire; Dify does not erase token bills or bad retrieval.
- License is not stock Apache — multi-tenant SaaS and logo stripping are the classic traps.
- Cloud vs Community feature and support boundaries change; verify current docs before assuming parity.
- Agent/workflow graphs fail the same way hand-rolled ones do: brittle tools, silent wrong answers, prompt injection — you still need eval.
- Default concurrency/worker knobs exist in compose configs; undersized boxes feel “broken” under load.
Who it's for
Teams prototyping LLM apps with a visual console who can accept the modified Apache terms (keep branding; don’t productize multi-tenant without a deal). Solo builders self-hosting a single workspace for internal tools. Not for people who need a permissive fork-and-white-label SaaS license without talking to LangGenius.
Related in this cluster
- AnythingLLM — MIT local-first RAG / agent UI
- n8n — fair-code workflow automation (broader than LLM apps)
- OpenHands — coding-agent platform
- LiteLLM — OpenAI-compatible gateway across providers
- Dev AI hub — more coding agents and local AI UIs
Verdict
Use Dify when you want a visual LLM app / RAG / agent platform and will either self-host Community Edition under the modified Apache terms or pay for Cloud/enterprise. Keep the logo if you ship the frontend, don’t stand up multi-tenant SaaS without a commercial path, and treat model bills and RAG eval as your problem — not the platform’s.