Dify: an LLM app development platform on a visual canvas

Moving an LLM app from demo to production rarely stalls on the model. It stalls on everything around it: how prompts get versioned, who builds the retrieval pipeline, how much code changes when you swap inference providers. Dify is one of the most-starred LLM app development platforms on GitHub (TypeScript, about 158.5k stars as of 2026-10-15): workflow, RAG, agents, model management and observability folded into one visual interface, with APIs facing outward.

Core features

• Visual workflow: build and test AI workflows on a canvas — nodes, branches and loops are dragged out, and every edit is testable immediately.
• Broad model support: hundreds of proprietary and open-source models from dozens of inference providers plus self-hosted solutions (GPT, Mistral, Llama3 and anything OpenAI-compatible); the provider list is maintained in the docs.
• Prompt IDE: craft prompts, compare model performance, and add text-to-speech to a chat app from one interface.
• RAG pipeline: the full path from document ingestion to retrieval, with out-of-the-box text extraction for PDFs, PPTs and other common formats.
• Agents with their own sandbox: the README describes agents that “run commands, install software, and handle files” to finish open-ended tasks. Give them skills, tools from Dify Marketplace, MCP servers or your own APIs, then run them as a chat app or as a step in any workflow.
• LLMOps and BaaS: improve prompts and datasets continuously from production data and annotations; every capability has a corresponding API to wire into your business logic.

Typical use cases

• A developer delivering a “turn the company knowledge base into an assistant” contract: ingestion, retrieval and prompts are configured in the platform, and what ships is a maintainable app instead of a pile of scripts.
• A product team comparing inference providers: the model layer is abstracted, so switching providers does not touch the application.
• A team that wants to self-host: Docker Compose to start, with official minimums of 2 CPU cores and 4 GiB of RAM.

Quick start

cd dify
cd docker
cp .env.example .env
docker compose up -d

Docker and Docker Compose v2.24.0+ are required. Once it is up, open http://localhost/install in a browser to finish initialization, then create your first app in the console. Advanced configuration, observability and deployment options live in docs/ADVANCED_SETUP.md.

Summary

For teams that need LLM apps to become maintainable products, and for developers delivering them — the platform absorbs the surrounding engineering (workflow, retrieval, model management) so the team can focus on business logic. It is not the tool for a one-off API call, where the infrastructure overhead does not pay off, nor for teams needing a deeply custom execution core: the platform’s abstraction is both convenience and boundary. As of 2026-10-15 the repo has 158,547 stars in TypeScript with pushes the same day — development is active. gh api reports the license as NOASSERTION, so check the LICENSE file and the official edition notes yourself (the Cloud, Community and Enterprise boundaries are set by the pricing page).

If all you need is a queryable memory layer for an agent rather than a whole platform, Graphiti builds a temporal knowledge graph where Dify manages the whole application path; and if you want an agent that writes code in a terminal, OpenClaw runs on your own devices.