Chatbots, agents and RAG apps, each with its own API
Dify is a web platform for building applications on top of LLMs: a chatbot that answers from a document base, an agent that calls tools, a workflow that chains models and services. Everything is set up in a console, workflows on a node canvas, and each published app gets its own API. The backend is Python (Flask, Celery) and the console is Next.js. You can self-host it or use Dify Cloud.
Flowise covers similar ground: agents and RAG chains wired together on a canvas, under Apache 2.0 except for its enterprise code. The project was archived in August 2026 and no longer ships fixes. n8n comes at it from the automation side: more than 1,500 service integrations, with AI features on top. Paperclip doesn't build applications at all: it coordinates agents you already run, such as Claude Code or Codex, through an org chart and budgets.
The comparison of open source workflow automation tools sets Dify's license and stack against the other tools.
Agents get their own sandbox, and RAG comes built in
- Workflows that start on their own: since 1.10, a workflow can fire on a schedule, on a webhook, or on a third-party event delivered by a plugin (Slack, GitHub, Linear).
- RAG built in: from document upload to retrieval, with text extraction from PDFs and slide decks. Weaviate is the default vector store, and the Compose file has profiles for sixteen others, Qdrant, pgvector, Milvus and Elasticsearch among them.
- Swappable models: hundreds of models from dozens of providers, plus anything that speaks the OpenAI API.
- Sandboxed agents: an agent runs commands, installs software and handles files in an environment of its own, locally or, since 1.17.0, on E2B. Its tools come from the Dify Marketplace, MCP servers or your own APIs.
- Production monitoring: logs and performance data in the console, plus Opik, Langfuse and Arize Phoenix integrations.
Single workspace for free, multi-tenancy by permission
The license is not plain Apache 2.0. Dify's LICENSE file adds two conditions. Running a multi-tenant environment, where one tenant equals one workspace, takes written permission from the publisher, LangGenius. Anyone using the web console (the web/ directory or the web Docker image) may not remove or alter Dify's logo and copyright notices. Commercial use is allowed, including as the backend for other applications. Contributors agree that the publisher can make the license stricter or looser, and GitHub doesn't match it to any known license.
Fifteen containers on startup. docker compose up -d brings up seven Dify services (API, websocket, worker, scheduler, console, plugin daemon, agent backend) and eight dependencies: PostgreSQL, Redis, Weaviate, Nginx, two Squid proxies that guard against SSRF, and two execution sandboxes. The docs ask for 2 CPU cores and 4 GiB of RAM, with 8 GiB for the Docker VM on macOS. MySQL can replace PostgreSQL.
The Community edition gets one workspace. Multiple workspaces, SAML or OIDC single sign-on, fine-grained RBAC, audit logs and an official Helm chart belong to the Enterprise edition, which is also self-hosted and priced on request. Outside Enterprise, Kubernetes deployments rely on community-written Helm charts that the repo lists.
Twenty-three stable releases in twelve months. The latest, 1.17.1, shipped on September 10, 2026. It moves the bundled Weaviate from 1.27.0 to 1.39.2: existing installs have to upgrade it by hand, in stages, without skipping a minor version, or vector search can break for good.