Dify

  • Dify — Full Profile ( 2026-09)

    LangGenius Inc. is a Shanghai-based and remote-first technology company founded in 2023 by Yuchen (Andy) Zhang and the Simon (Zhang) core team. In March 2026, the firm closed a $30M Series Pre-A financing round led by HSG, with participation from GL Ventures, Alt-Alpha, 5Y, Mizuho and NYX investors. While positioning as VC-independent leaning in line with n8n, Databox and Apify, LangGenius stands out with substantially heavier institutional funding compared with other vendors in the independent cohort. As of August 2026, its flagship open-source project langgenius/dify has accumulated 150K–153K GitHub stars, serving more than 280 enterprise clients including Maersk, ETS, Novartis, Volvo and Anker, with over 1.4 million deployed machine instances powering global enterprise workflows. Starting from v1.6.0 released in July 2025, Dify shipped industry-first bidirectional first-party MCP capabilities, enabling a unique dual operational role in the MCP ecosystem. It functions both as an MCP client capable of invoking external MCP servers such as Linear, Notion and Zapier, and as an MCP server by exposing native Dify workflows via public /mcp endpoints for consumption by agent clients including Claude and Cursor. Ecosystem-wise, LangGenius joins the stable first-party GA server cluster alongside Klaviyo, Feedly, Slack, Databox, Tavily and Apify. Its bidirectional client-server MCP model is most analogous to n8n’s dual-GA architecture, with clear differentiation: n8n delivers general-purpose stable dual GA capability, while Dify’s bidirectional implementation is strictly application-centric. On licensing, Dify adopts a custom modified Apache-2.0 “Dify Open Source License”, explicitly excluding standard Apache-2.0 and AGPL-3.0 terms, creating a distinct open-source legal profile separate from AGPL-licensed vendors including Documenso and QuickChart.

    Website

    Field Value
    Official Website https://dify.ai · pricing: dify.ai/pricing · docs: docs.dify.ai · MCP docs: docs.dify.ai/features/agent/mcp · GitHub: github.com/langgenius/dify (153K★ Aug 2026, modified Apache-2.0) · self-host: Docker Compose / Helm (github.com/langgenius/dify/blob/main/docker)
    Company LangGenius Inc., Shanghai + global remote; founded 2023; $30M Pre-A Mar 2026 HSG-led; 150K+ stars, 280+ enterprise customers, 1.4M machines; modified Apache-2.0 (Dify OSL: no multi-tenant SaaS w/o commercial license, no logo strip); Cloud-hosted + self-host Docker/K8s/VPC; SOC 2 Type II + ISO 27001 on Enterprise, SSO/SAML/RBAC/audit-logs, GDPR-aware ToS
    Status Active, v1.17.0 Aug 2026 (E2B sandbox, skill management, Agent open beta), MCP bidirectional GA since v1.6.0
    Category ai-productivity-data
    Pipeline Stage none

    Title

    Field Value
    Product Dify (LLM app platform: visual workflow + RAG + agent runtime + model gateway + MCP bidirectional + self-host)
    Descriptive Title Dify — Open-Source LLM Application Platform for Production-Ready Agentic Workflows: drag-canvas orchestration (branching/loops/HTTP/code/human-in-loop nodes) + RAG pipeline (ingest→chunk→embed→hybrid-retrieve→rerank) + agent runtime (Function-Calling/ReAct, 50+ built-in tools) + model gateway (OpenAI/Anthropic/Gemini/DeepSeek/Qwen/GPT/Ollama/Xinference, OpenAI-compatible) + observability (logs/traces/Langfuse-Langsmith-Arize) + every app publishes as REST API / chat widget / web app / MCP server URL; MCP bidirectional GA v1.6.0​ — Dify agents call external MCP servers (Linear/Notion/Zapier via HTTP MCP 2025-03-26) AND any Dify workflow/agent exposes /mcp for Claude/Cursor/ChatGPT; Cloud Sandbox free 200 credits → Professional 159/mo (10K/50 seats) → Enterprise custom (VPC/Helm/SSO/SLA); self-host Community Edition 30M-Pre-A-HSG 150K★ MCP-bidirectional-GA-not-plain-apache-not-agpl
    One-line Positioning
    In practical workflows, users visually build RAG and agent-based workflows on the Dify canvas and publish the customized pipeline as a dedicated /mcp endpoint. External AI clients such as Claude and Cursor can mount this endpoint and treat the custom Dify agent workflow as a native tool. Meanwhile, Dify’s built-in agent runtime can itself call external MCP servers, including Notion, enabling cross-platform tool orchestration.
    This gives Dify unique bidirectional MCP functionality: it operates both as an MCP server that exposes self-built AI tools externally and as an MCP client that consumes third-party MCP services. This distinguishes it from single-direction MCP vendors such as Tavily and Apify, which only serve as MCP servers and cannot initiate outbound MCP calls. While n8n also supports dual client/server MCP behavior, n8n is fundamentally general-purpose and automation-centric, whereas Dify is LLM-application-centric — built specifically for AI agent deployment rather than generic protocol-level automation.
    Dify adopts a hybrid billing model. Its cloud offering monetizes through workspace tiers, seat subscriptions, and additive message credits. Self-hosted deployments remain completely free under Dify’s modified Apache open-source license. This pricing structure creates clear reverse differentiation against competitors: it differs from Motion’s seat-only model, Tavily’s compute-unit billing, Databox’s AI credit metering, and QuickChart’s image-count based pricing.
    In terms of company ownership, Dify is a VC-backed private company, placing it in the same independent commercial vendor bucket as n8n, Databox, Apify, Typeform, and Documenso. It is structurally distinct from acquired-but-independent platforms like Tavily, nonprofit public-good entities such as Zotero-CDS, and fully bootstrapped tools like QuickChart.

    Features

    Field Value
    Workflow Canvas Visual DAG: LLM/agent/code/HTTP/IF-else/loop/variable/question-classifier/template/knowledge-retrieval/human-input/end nodes; branching + parallel; version history; diff between versions; import/export DSL JSON
    RAG Pipeline Document ingest (PDF/MD/DOCX/PPTX/HTML/notion/website), chunk strategies (general/parent-child/qa), embedding providers (OpenAI/Cohere/Jina/local), hybrid BM25+vector, rerank (Cohere/Jina/custom), metadata filter, 7K-CSL not relevant but dataset annotation loop
    Agent Runtime Agent strategy Function-Calling + ReAct, 50+ built-in tools (web-search/code-exec/DALL·E/serp/HTTP), custom tool via OpenAPI spec or Python snippet, multi-agent delegation (sub-agent node), Agent mode entered open-beta Jul 2026 with skill management
    Model Gateway 100s of models: GPT-4o/Claude/Gemini/DeepSeek/Qwen/Mistral/Llama/Ollama/Xinference/vLLM, OpenAI-compatible endpoint passthrough, per-app model routing, fallback chain, token accounting per provider
    MCP (bidirectional GA v1.6.0) As client: add MCP server URL (HTTP/SSE/streamable) → tools auto-discovered into agent tool palette (Linear/Notion/Zapier/Tavily/Apify all demonstrable); protocol 2025-03-26; As server: any published Dify app/agent/workflow gets /mcp endpoint (Bearer or session token) → Claude/Cursor/ChatGPT/5ire mount it as resource+tool; not a separate mcp.dify.com hosted like Tavily/Apify, it’s per-deployment URL → first-party-but-app-scoped (between n8n dual-GA and Klaviyo server-only)
    Publish Surface OpenAI-compatible REST /v1/chat-messages etc, chat widget embed, web app with auth, API key scoping, rate-limit, annotation feedback loop
    Observability Run logs, token cost per model, trace spans, Langfuse/Langsmith/Arize Phoenix export, prompt A/B, user-session replay
    Enterprise SSO/SAML, RBAC, audit logs, SOC2+ISO, Helm K8s, VPC, Azure/AWS Marketplace, 24/7 CSM
    License Modified Apache-2.0 (Dify OSL): self-host internal OK, build/sell your own app on Dify OK, no multi-tenant Dify-SaaS resell w/o LangGenius written license, no logo strip in web console; not OSI-pure, not AGPL (contrast Documenso/QuickChart AGPL-3.0)

    Pricing (2026-09, cloud workspace+credits+seats / self-host $0 modified-Apache)

    Field Value
    Sandbox (Cloud) $0​ — 200 message credits lifetime, 1 workspace, 1 member, 5 apps, 50 knowledge docs, 50MB storage, 3K trigger events, 30-day logs
    Professional $590/yr) per workspace — 5,000 msg credits/mo, 3 seats, 50 apps, 500 docs, 5GB, priority doc processing, 20K triggers, unlimited logs, no API rate limit
    Team 1,590/yr) per workspace — 10,000 msg credits/mo, 50 seats, 200 apps, 1,000 docs, 20GB, unlimited triggers, top-priority workflow queue, 5K annotation quota
    Enterprise Custom — VPC/Helm, SSO/SAML, audit, SOC2 reports, DPA, SLA, >50 seats, custom credits
    Self-Hosted Community $0​ — Docker Compose / K8s, full feature parity with cloud minus managed uptime, modified Apache-2.0, LLM inference billed to your own provider key (OpenAI/Anthropic/etc) separately​ — Dify credit only meters platform features not tokens
    Message Credit Covers hosted-model trial calls (OpenAI/Anthropic/Gemini/xAI/Qwen via Dify key); once exhausted switch to BYO key; credit≠token, 3-5x naive forecast needed
    Type workspace-seat-credit-hybrid-no-per-call; reverse Motion pure-per-seat, reverse Qualtrics per-interaction, reverse Tavily CU, reverse Databox AI-credit, reverse QuickChart image-count, reverse Gmail 80M-unit, reverse Semantic Scholar 1-RPS, reverse Zotero storage, reverse Apify CU — Dify is seat+credit+workspace​ triple gate
    Gotcha Multi-workspace = multiply 159; 200-credit sandbox not real eval; self-host needs 2CPU/4GB min + Postgres+Redis+vector+ sandbox container; modified-Apache multi-tenant resell = license violation

    Reviews

    Field Value
    G2 4.6/5 ~2,221 reviews​ (RFP.wiki aggregate Jul 2026) / 4.1/5 ~20-40 verified​ (direct G2 scrape) → print G2 4.6/~2221-aggregate OR 4.1/~20-40-verified (volume real, devtool-B2B solid, thinner-sample direct than aggregate; aligns Klaviyo/Databox/Apify vc-SaaS G2-rated pattern, not Tavily 2-review)
    Capterra 4.5/5 ~308 reviews​ (RFP.wiki) / 4.2-4.6 ~20-40​ (direct) → Capterra 4.5/~308-aggregate directional-solid
    Trustpilot 1.2/5 ~1,035 reviews​ — but these are Dify.ai consumer-chatbot complaints not platform-API​ (mixed B2C confusion) → print Trustpilot 1.2/~1035 B2C-confused-excluded (aligns FB-Ads/Gmail consumer-score exclusion rule, do NOT average into 4.x)
    Software Advice 4.5/~306 directional
    Dev / GitHub 153K★ langgenius/dify, 22K forks, 800+ contributors, v1.17.0 Aug 2026 active; AIMultiple/Context.dev 2026 “category leader by stars”, gripes credit opacity + canvas ceiling + multi-tenant license surprise + self-host ops weight
    Praise Most-complete OSS LLM-app stack (workflow+RAG+agent+MCP+observe), 150K★ momentum, bidirectional MCP (rare: both client+server), BYO-model cost control, Maersk/Novartis proof, modified-Apache still self-host-free
    Gripes Message-credit opacity (3-5x forecast), canvas hits ceiling on complex logic (drop to code), multi-tenant license clause surprises SaaS-wannabes, self-host = Postgres+Redis+vector+sandbox heavy, cloud per-workspace pricing multiplies fast, Trustpilot flooded by B2C chatbot-misuse 1-star

    Review Sources Breakdown

    Field Value
    G2.com 4.6/5 ~2,221 (aggregate RFP.wiki Jul 2026) / 4.1/5 ~20-40 (direct verified) → G2-solid-vc-saas-pattern
    Capterra 4.5/5 ~308 aggregate / 4.2-4.6 direct → Capterra-solid
    Trustpilot 1.2/5 ~1,035 but B2C chatbot-misuse not platform → no-rated-platform-profile-excluded-consumer-confused
    Software Advice 4.5/~306 → directional
    GitHub 153K★ + AIMultiple/Context.dev dev-sentiment directional
    Conclusion​ Prints “G2 4.6/~2221-agg (4.1/~20-40 direct) + Capterra 4.5/~308 + TP 1.2/~1035 B2C-excluded + GitHub 153K + dev consensus category-leader-by-stars-credit-opaque-canvas-ceiling” — aligns vc-independent SaaS G2-rated (Klaviyo/Databox/Apify), excludes TP like FB-Ads/Gmail.

    Access

    Field Value
    Web App cloud.dify.ai (managed) or self-host :3000 install wizard
    API REST OpenAI-compatible /v1/chat-messages, /v1/completion-messages, /v1/workflows/run, /v1/datasets, Bearer APP-KEY; SDK JS/Py; LangChain DifyAPI wrapper
    MCP First-party bidirectional GA v1.6.0​ — (1) Settings→MCP Servers add external URL (HTTP/SSE/streamable 2025-03-26) → tools appear in agent node; (2) any published app → https://<your-dify>/apps/<id>/mcp (Bearer or session) → Claude/Cursor/ChatGPT mount; not centralized mcp.dify.com, per-deployment URL (app-scoped first-party, between n8n dual and Klaviyo server-only)
    SDKs dify-client-js, dify-client-python, LangChain Dify APIChain, CrewAI tool, n8n HTTP node, Zapier/Make connector
    Login email/GitHub/Google/OAuth; Enterprise OIDC/SAML
    Self-host Core differentiator​ — Docker Compose (2CPU/4GB min), Helm K8s, Postgres+Redis+vector (Weaviate/Qdrant/Milvus/PgVector)+sandbox container; modified Apache-2.0; multi-tenant resell needs commercial license

    Best For

    Field Value
    Audience 1 Teams building internal RAG chatbots / doc-QA on own data with visual canvas (no LangChain-from-scratch)
    Audience 2 Startups shipping LLM app to API + web widget fast, BYO OpenAI/Anthropic key, self-host for data residency
    Audience 3 Agent builders wanting Dify-as-MCP-server (expose workflow to Claude) AND Dify-calls-Notion/Linear-MCP-client in same canvas
    Audience 4 Enterprises needing SSO/SAML/RBAC/audit/SOC2 on a self-hosted LLM app layer (not just model gateway)
    Audience 5 Prototypers using Sandbox 200 credits to validate before $59 Pro
    Audience 6 Multi-model routing teams (DeepSeek+Claude+GPT fallback chain in one node)

    Not For

    Field Value
    Exclude 1 Teams wanting plain Apache-2.0/AGPL to resell as multi-tenant AI-SaaS — Dify OSL forbids w/o commercial license (use Flowise/Langflow/Crawlee)
    Exclude 2 Non-technical no-code-only users — canvas needs prompt/retrieval/var comfort (use Coze/Gumloop)
    Exclude 3 Extreme agent logic beyond canvas (custom RL/stateful graph) — drop to LangGraph/LangChain code
    Exclude 4 Buyers refusing credit-metered cloud math — self-host $0 but owns Postgres+Redis+vector+sandbox ops
    Exclude 5 Pure model-gateway-only (no workflow/RAG UI) — LiteLLM better fit
    Exclude 6 Teams needing first-party centralized mcp.dify.com hosted endpoint — Dify MCP is per-deployment URL, not Tavily/Apify-style hosted registry

    Competitors

    Field Value
    Flowise Apache-2.0 drag LLM flow, lighter than Dify, no RAG-admin depth, 30K★, no multi-tenant license claw, weaker enterprise
    Langflow Apache-2.0 (acquired by DataStax), visual LangChain wrapper, 35K★, flow-centric not app-platform, no MCP-server-publish maturity
    Coze (ByteDance) No-code bot builder, closed SaaS, non-tech users, no self-host, no MCP-server-publish (client only)
    RAGFlow AGPL-3.0, document-deep RAG engine, 80K★, weaker workflow/agent canvas, no MCP bidirectional GA
    n8n Fair-code automation with AI nodes, dual-MCP GA, but not LLM-app-platform (no RAG-admin/chat-widget/publish-as-API native)
    LangChain/LangGraph Code-first, no UI/canvas, Dify often wraps or complements
    Botpress Commercial bot platform, $495/mo floor, weaker RAG-openness

     

Do Not Sell or Share My Personal Information Cookie Settings