Dify (GitHub: langgenius/dify) is an open-source platform for building and operating AI applications — visual workflow orchestration, RAG pipelines, agent execution, and API publishing in one self-hostable package. In March 2026, the company (operated by LangGenius, Inc.) closed a $30 million Series Pre-A led by HSG, with GL Ventures, Alt-Alpha Capital, 5Y Capital, Mizuho Leaguer Investment, and NYX Ventures participating. At the time of that announcement, Dify described itself as the 51st most-starred open source project on GitHub, running on more than 1.4 million machines worldwide across 175+ countries.
The headline metric: 150,000+ GitHub stars and 280+ enterprise customers — including Maersk, ETS, Anker, and Novartis — with over 1.4 million machines running Dify worldwide, as of mid-2026. That is not a prototype-friendly toy — it is a production platform with real adoption.
What Dify Actually Does
Dify sits at the intersection of several categories that are usually separate tools:
- Visual workflow builder: drag-and-drop orchestration of LLM calls, conditionals, loops, retrieval steps, and human-in-the-loop nodes
- RAG pipeline: document ingestion, chunking, embedding, and retrieval across multiple knowledge bases — widely regarded as its strongest feature
- Agent runtime: supports Function Calling and ReAct strategies, with 50+ built-in tools (web search, code execution, image generation) and custom tool definitions
- Model management: one interface for “seamless integration with hundreds of proprietary/open-source LLMs from dozens of inference providers and self-hosted solutions”, covering GPT, Claude, Gemini, DeepSeek, Qwen, Llama, Mistral, and any OpenAI-compatible endpoint
- API publishing: every Dify application exports as an API endpoint, a chat widget, or a web app with built-in authentication and rate limiting
- Observability: full tracing, prompt version history, and user analytics built in
The self-hosted Community Edition (Docker Compose, single machine or Kubernetes) is free and includes the core platform, but it ships under a source-available license, not plain Apache/MIT: self-hosting for your own org and building commercial products with the API are permitted, but operating Dify itself as a multi-tenant SaaS product requires a commercial license from LangGenius. Dify Cloud starts with a free Sandbox tier and scales to Professional ($59/month or $590/year) and Team ($159/month or $1,590/year), with Enterprise priced at $150,000 for a 12-month license on AWS Marketplace (custom pricing also available via private offer).
2026 Updates: What Changed
Two capabilities added in 2026 matter for the builder decision:
MCP as Client and Server
As of Dify v1.6.0 (July 2025), Dify supports the Model Context Protocol bidirectionally.
As an MCP client, agents inside Dify can call any external MCP server directly — the release notes cite examples like Linear, Notion, and Zapier — using HTTP-based MCP (protocol version 2025-03-26). This eliminates per-service integrations in favor of a single discovery mechanism.
As an MCP server, you can expose any Dify agent or workflow as an MCP server for other clients: after documenting the service and its parameters, Dify issues a standard MCP server URL, making the workflow callable from Claude, Cursor, or any other MCP client. The workflow-to-MCP path is particularly valuable for teams that want visual tooling for building but programmatic access for consuming.
Human Input Node
Shipped in Dify v1.13.0, the Human Input node lets a running workflow pause for human review and resume with approved, edited, or rerouted decisions. Reviewers can respond via web app or email, edit variables before the workflow continues, and route down different branches via configurable buttons (e.g., “Approve,” “Reject,” “Escalate”); a configurable timeout (3 days by default) sends unanswered requests down a fallback branch. This is more structured than ad-hoc tool-call interrupts; you define pause points at design time.
For regulated industries and high-stakes workflows (legal review, compliance checks, medical triage), this node is the feature that makes Dify viable where fully autonomous agents are not.
Strengths
RAG pipeline depth. Dify’s knowledge base capabilities are production-grade: multiple retrieval strategies (similarity, full-text, hybrid), configurable chunk sizes, custom embedding models, and retrieval scoring. Most competitors offer RAG as a bolted-on feature; Dify was built with document retrieval as a first-class citizen.
Self-hosted by default. For teams with data residency requirements, enterprise compliance obligations, or a preference to keep inference traffic off third-party infrastructure, the self-hosted path is the primary option — not an afterthought. The Docker Compose setup deploys the full stack including worker queue, a configurable vector DB (Qdrant, Weaviate, Milvus, PGVector, and several others are supported as Docker profiles), and cache.
Model-agnostic. Switching the underlying LLM in a Dify workflow is a configuration change, not a code change. This is practical insurance against vendor lock-in and cost optimization — if GPT-5.6 ships at a better price-performance ratio than your current model, swapping it into a Dify workflow doesn’t require touching the orchestration logic.
Visual debugging. Workflow runs are traceable step by step. You can inspect inputs and outputs at every node, replay steps with modified inputs, and compare prompt versions against each other. This is materially better than debugging LangGraph’s Python graphs through log files.
Weaknesses
GUI ceilings. The visual canvas is productive for common patterns and degrades for uncommon ones. Complex branching logic, custom retry strategies, and anything requiring dynamic code generation against schema produces either deeply nested nodes or workarounds that would have been simpler in code. Dify is not a replacement for LangGraph or the Claude Agent SDK in applications where the control logic is the product.
Node graph performance. Large workflows with many parallel branches and complex retrieval pipelines can become slow to render and navigate in the browser canvas. This is a tooling problem, not a runtime problem, but it affects developer velocity at scale.
Ecosystem maturity gap. LangChain, LangGraph, and LlamaIndex have extensive third-party tutorials, integration libraries, and StackOverflow coverage. Dify’s documentation is good, but the surrounding community is thinner. You are more likely to hit an undocumented edge case and need to file a GitHub issue than with more established frameworks.
Cloud pricing mismatch for small teams. The Professional tier ($59/month) and Team tier ($159/month) are reasonable for organizations, but the jump from the free Sandbox to the first paid Cloud tier may feel abrupt for individual builders. Self-hosting resolves this, but requires Docker familiarity and server maintenance.
Competitive Position
| Dimension | Dify | LangGraph | Flowise | n8n |
|---|---|---|---|---|
| Design surface | Visual canvas | Python/TypeScript code | Visual canvas | Visual canvas |
| RAG | Best-in-class | Framework only | Good | Plugin-based |
| Production readiness | High | High (code) | Medium | High (ops) |
| MCP support | Client + Server | Client (with adapters) | Limited | Limited |
| Self-hosted | Core offering | Core offering | Core offering | Core offering |
| AI focus | AI-first | AI-first | AI-first | Mixed (AI + ops) |
Dify wins when the team needs visual orchestration with strong RAG and doesn’t want to write glue code. LangGraph wins when the control logic requires code expressiveness and the team is comfortable in Python. Flowise wins for simpler chatbot-over-documents use cases with minimal setup. n8n wins when AI is one step in a larger business automation that also connects Salesforce, HubSpot, and a payment processor.
Who Should Use Dify
Appropriate for:
- Teams building document-heavy applications (contract review, research assistants, support bots) where RAG depth matters
- Organizations with data residency requirements that want self-hosted inference routing
- Builders who need visual tooling for non-engineer stakeholders to modify workflows without code changes
- Enterprises needing human-in-the-loop approval nodes in otherwise automated pipelines
- Teams building MCP-accessible tools: Dify’s workflow-to-MCP server path is fast and does not require writing an MCP server manually
Not appropriate for:
- Applications where the agent’s control logic is highly custom (complex dynamic planning, recursive self-modification, multi-model deliberation with shared state)
- Teams running high-throughput real-time inference — Dify’s overhead is acceptable for many workloads but adds latency relative to direct API calls
- Projects where LangGraph or the Claude Agent SDK is already working well at scale
Rating: 4/5
Dify is production-ready for the workflows it is designed for. The RAG pipeline is among the best available in an open-source package. The 2026 updates — MCP bidirectionality and the Human Input node — bring it current with the state of the field.
The limitations are real: visual tooling has a ceiling, and complex custom logic belongs in code. But for the large category of AI applications that involve documents, knowledge retrieval, structured workflows, and human review checkpoints, Dify removes weeks of integration work that would otherwise fall on the engineering team.
The self-hosted path, the model-agnostic design, and the active 150K+-star community make it a credible long-term foundation — not a vendor bet that creates lock-in.