// the find
langgenius/dify
Production-ready platform for agentic workflow development.
Dify is a self-hostable platform for building LLM applications — think visual workflow builder, RAG pipelines, and agent orchestration wrapped in a web UI. It targets teams that want to ship AI-powered apps without writing orchestration code from scratch. The audience is product teams and ML engineers who want a control plane, not researchers who want flexibility.
- Model provider breadth is real and maintained — 50+ providers with consistent abstraction, not just OpenAI wrappers with an adapter bolted on.
- The workflow canvas handles non-trivial DAGs: branching, iteration, parallel execution, and human-in-the-loop nodes are first-class, not afterthoughts.
- Docker Compose cold start is genuinely straightforward — single `.env` copy and `compose up`, no multi-step config dance.
- LLMOps surface (token usage, latency traces, annotation-based dataset improvement) is production-grade and built in, which saves integrating a separate observability layer.
- The Python backend is Flask + Celery — fine for low concurrency, but the architecture does not scale horizontally without careful Redis and worker tuning that the docs underexplain.
- The visual workflow editor leaks abstraction constantly: complex graphs become unreadable JSON DSL that you end up editing by hand, and there is no meaningful diffing or version control story beyond 'export and compare'.
- Plugin/tool development is poorly documented; the extension points exist but the contracts are underdocumented, so third-party integrations frequently break on minor version bumps.
- The enterprise vs. community feature split is opaque — several features visible in the UI or docs are silently gated behind a license, and you only discover this after building a workflow that depends on them.