Dify

Dify is an open-source AI agent builder where developers and teams design agentic workflows and RAG apps on a visual canvas, run in the cloud or self-hosted.

Dify turns a model, a set of documents and a few API tools into an AI application that runs in production. Builders lay the logic out as nodes instead of writing orchestration code, test each step and publish the result. LangGenius, Inc., an independent Delaware corporation, develops it in the open on GitHub. It is offered as Dify Cloud, as an Enterprise private deployment and as an open-source edition that teams run themselves. Workflow and Chatflow apps share one canvas, where an Agent node calls models and tools as one step beside LLM and Knowledge Retrieval nodes. The Knowledge Pipeline turns uploaded or fetched documents into an indexed knowledge base that apps retrieve from at run time. Tools come from OpenAPI specifications imported as custom tools or from any connected MCP server. A finished app can also be published as an MCP server. The Human Input node pauses a run for a person to approve, while Webhook, Plugin and Schedule triggers start Workflow apps from outside events. Run logs trace every node with its inputs, outputs and token use. Version control keeps the draft apart from the published app, so an earlier version can be restored. A standalone Agent builder, still in beta, pairs a chosen chat model with tools, custom APIs and workflows. Dify suits independent developers and small product teams putting a first AI feature into production. Medium-sized teams use it when several builders work on the same apps. Enterprises choose it when models and data must stay inside their own network. Dify Cloud is bought per workspace, monthly or annually, and each plan sets a budget of message credits and caps on members and apps. Enterprise deployments are quoted by sales, and the open-source edition runs through Docker Compose on a laptop, a server or a cloud machine. Support and internal help teams answer questions from their own documents with it. The assistant sits on a website as a chat bubble or inside an internal tool through an iframe. Operations teams build event-driven automations, such as a workflow that starts on a GitHub event or a new Gmail message. That workflow can wait for a person's approval before it acts. Product engineers call published agents from their apps through the service API and follow each run in Langfuse or LangSmith. The usual path runs from a prototype on the canvas and a test run on real documents to an endpoint the product already calls.