Dify
Also known as: LangGenius
Open-source LLM app development platform for building agents, chatbots, and workflows.
Dify is an open-source platform for building LLM applications, designed to bundle the many moving parts of an AI product into one coherent stack. Building on top of large language models normally means stitching together model access, prompt engineering, retrieval, agent logic, and monitoring from a dozen separate tools. Dify replaces that fragmented toolchain with a single self-hostable environment, accessible through an intuitive web interface, so teams can go from prototype to production without writing the underlying orchestration plumbing.
At its center is a visual workflow builder where you assemble multi-step AI applications on a canvas, combining language models, web search, code execution, and your own APIs with branching logic. Alongside it sits a Prompt IDE for crafting and comparing prompts across models, and a built-in RAG pipeline that handles the whole retrieval flow: ingesting documents like PDFs and slides, chunking and embedding them, storing them in a vector database, and retrieving and reranking the right context at query time. This Knowledge surface is a first-class part of the product, giving teams a concrete place to manage data and grounding rather than rebuilding retrieval in every app.
Dify is firmly model-agnostic. It connects to hundreds of proprietary and open-source models across many inference providers, from commercial APIs to locally hosted models, and lets you switch the model behind an application without changing code. Its agent capabilities let you define agents using function calling or a reasoning-and-acting loop, equip them with dozens of built-in tools or your own, and connect external tools through the Model Context Protocol, with the option to turn a finished Dify workflow into an MCP server itself.
Rounding out the platform are LLMOps features for monitoring logs and performance and improving prompts and datasets from real usage, native integrations with observability tools, and a backend-as-a-service layer that exposes every app as an API for embedding into existing software. Dify can be run as a managed cloud service or self-hosted for full data control, which makes it a common choice for teams with privacy, compliance, or requirements who still want a fast path to shipping chatbots, copilots, and agents.
Vendor details
Canonical URL
https://dify.ai
Category
Agent builder
Subcategory
Open-source visual LLM app and agentic workflow platform (Workflow Studio, Knowledge Pipeline, Marketplace)
Company status
independent
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
Integrations connect Dify to model providers, tools, data sources and external services, installed as plugins from a public marketplace typed across Models, Tools, Data Sources, Triggers, Agent Strategies, Extensions and Bundles. In-workflow surfaces are a Tool node, an HTTP Request node for arbitrary APIs, MCP tools, Custom Endpoints, an External Data Tool extension and a Sensitive Content Moderation endpoint. Credit-metered hosted models named on the pricing page are OpenAI, Anthropic, Gemini, xAI and Tongyi, and customers can connect their own provider accounts and keys or, in air-gapped deployments, locally deployed models. Knowledge ingestion covers local file upload, Notion sync and website crawl, plus an External Knowledge API specification for pointing Dify at an existing RAG service. Outbound, a published workflow is exposed as a hosted web app, an iframe or chat-widget embed, a REST API with a published OpenAPI specification, the difyctl CLI, or an MCP server consumable by Claude Desktop and Cursor. Traces stream to Langfuse, LangSmith, Opik, W&B Weave, Arize, Phoenix and Alibaba Cloud ARMS over OpenTelemetry.
In practice
You want a support chatbot grounded in your own documentation without assembling a RAG stack yourself. Dify's built-in Knowledge pipeline ingests, chunks, embeds, and retrieves your docs, and you wire it into a workflow visually.
Your team is stitching together model access, prompts, retrieval, and monitoring from separate tools. Dify combines workflow building, a Prompt IDE, RAG, agents, and observability in one self-hostable platform from prototype to production.
You need to keep data on your own infrastructure and stay free to switch models. Dify is open source and self-hostable, connects to hundreds of commercial and local models, and exposes every app as an API.
Sources & related URLs
Related / legacy domains
Research sources
Agentic Index coverage score
12.0 / 14 capabilities · 86%
| Integrations & Tool Calling | Full |
|---|---|
|
Integrations are documented as connecting Dify to model providers, tools, data sources and external services, with a Tool node, a Dify Tools management surface, an HTTP Request node for arbitrary APIs, MCP tools, Custom Endpoints, an External Data Tool extension and a public marketplace of installable plugins typed across Models, Tools, Data Sources, Triggers, Agent Strategies, Extensions and Bundles. Agent nodes invoke tools autonomously with per-parameter control over whether values are model-generated or fixed at configuration time. Sourcedocs.dify.ai integrations and node reference, marketplace.dify.airead 2026-08-31 |
|
| Workflow Orchestration | Full |
|
Workflow and Chatflow share one node library and execution model, with documented If-Else branching, Iteration over arrays, Loop with progressive refinement, Question Classifier routing, Variable Aggregator convergence, Parameter Extractor, sandboxed Code execution, and Orchestration Logic covering how nodes are arranged, nested or reused. Agent nodes run a reasoning-and-tool loop as a step inside a larger flow with a configurable iteration ceiling, a published workflow can be called as a tool by another workflow or exposed as an MCP server, and version control plus Snippets support reuse across flows. Sourcedocs.dify.ai workflow, node and version-control referencesread 2026-08-31 |
|
| Knowledge Grounding & RAG | Full |
|
Knowledge is a first-class product surface with its own documentation tree: ready-to-use knowledge bases with configurable chunking, cleaning and indexing method, a custom Knowledge Pipeline built and published as its own orchestrated flow, ingestion from local files, Notion sync and website crawl, document metadata management, a Knowledge Retrieval node, a retrieval hit-testing tool, and an External Knowledge API specification for pointing Dify at a customer's existing RAG service. Storage and request-rate quotas are metered per tier on the pricing page. Sourcedocs.dify.ai knowledge tree and dify.ai/pricingread 2026-08-31 |
|
| Human Oversight & Guardrails | Full |
|
Human Input is a first-class workflow node documented in the node reference, described as pausing the workflow to request human input; the Workflow Studio product page describes the same mechanism as pausing before actions that affect sensitive data, access or policy and letting a person approve, edit, comment or forward the run before it continues, with timeout handling. A Sensitive Content Moderation endpoint extension is a separate documented guardrail, and workspace member roles govern who can act. Sourcedocs.dify.ai node reference and dify.ai/workflowsread 2026-08-31 |
|
| Security, Identity & Governance | Full |
|
SOC 2 Type I, SOC 2 Type II, ISO 27001:2022 and GDPR are stated in the vendor's own compliance-report policy page, with the ISO assessment attributed to auditing firm Johanson and reports released to paid customers on request via security@dify.ai. A trust center resolves at security.dify.ai with a subprocessors page, and a Data Protection Agreement is published as a signed PDF alongside the privacy policy, terms and marketplace agreement. Named customer-facing controls: SSO via SAML/OIDC, fine-grained RBAC down to workflow level, SCIM provisioning, session policies, tamper-evident audit logs streamed to the customer's SIEM, PII-redacted prompt history and bring-your-own-key encryption; SSO and the enterprise control set are gated to the Enterprise tier. Sourcedocs.dify.ai/en/policies/agreement/get-compliance-report, dify.ai/dify-enterprise and security.dify.airead 2026-08-31 |
|
| Observability & Auditability | Full |
|
Monitoring is documented as a built-in analytics dashboard for performance, cost and engagement, plus conversation and run logs with stated retention and downloadable archived workflow logs, a Run History view, a Variable Inspector for intermediate values, and typed error metadata per node. Traces stream to seven named external stacks including Langfuse, LangSmith, Opik, W&B Weave, Arize, Phoenix and Alibaba Cloud ARMS over OpenTelemetry. On Enterprise the vendor names tamper-evident audit logs streamed to the customer's SIEM, per-call traces and PII-redacted prompt history. Sourcedocs.dify.ai monitor tree and dify.ai/dify-enterpriseread 2026-08-31 |
|
| Memory & State Persistence | Partial |
|
State persists within a conversation but not across them. A Variable Assigner node manages persistent conversation variables in Chatflow applications, Chat Web Apps keep persistent history, and the Agent node exposes a Memory toggle with a configurable window size implemented as a token buffer over recent messages. Log history is retained and browsable per tier, 30 days on Sandbox and unlimited on the paid plans, and workflow runs can be inspected after the fact. No cross-session or learned memory feature appears anywhere in the 265-page cloud documentation index. Sourcedocs.dify.ai node reference and dify.ai/pricingread 2026-08-31 |
|
| Deployment & Data Residency | Full |
|
Four documented topologies: hosted Dify Cloud, a Dify-managed single-tenant dedicated cluster in the customer's region, customer-managed deployment into the customer's own AWS, GCP or Azure VPC via Helm chart or Terraform modules, and fully air-gapped on-prem running locally deployed models with no egress. The open-source Community Edition self-deploys with Docker Compose and carries its own 277-page self-hosting documentation tree. Data residency is stated as keeping data inside the customer's VPC, region and perimeter, with bring-your-own-key encryption and multi-tenant isolation offering tenant-level quotas and access policies. Sourcedify.ai/dify-enterprise and docs.dify.ai self-host treeread 2026-08-31 |
|
| Prebuilt Agents, Templates & Packs | Full |
|
Dify Marketplace carries a named templates catalog of ready-to-use workflows that open in Dify with one click, filterable by Marketing, Sales, Support, Operations, IT and Knowledge and by partner, with entries such as Daily AI News Digest by langgenius, Competitive Landscape Intelligence, AI Technical Research Assistant and a Qdrant upsert on a Google Drive trigger, each showing its publisher and usage count, alongside a plugin catalog typed across Models, Tools, Data Sources, Triggers, Agent Strategies, Extensions and Bundles and a Creator Center. A documented Publish Apps to Marketplace flow lets customers publish their own apps, workflows export and import as DSL, and Snippets reuse groups of nodes. Sourcemarketplace.dify.ai/templates and docs.dify.ai publish-to-marketplaceread 2026-09-29 |
|
| Triggers & Channel Coverage | Full |
|
The node reference carries a dedicated Trigger family with three documented members, Schedule Trigger, Integration (plugin) Trigger and Webhook Trigger, and the Start node is documented as choosing whether a workflow begins on demand or automatically from a user message, API call, scheduled trigger, webhook, plugin event or uploaded payload. Channels on the output side are a hosted web app, an iframe or chat-widget embed, the REST API, the difyctl CLI and a published MCP server. The pricing page corroborates the feature as shipped rather than announced by metering Trigger Events across all tiers, 3,000 on Sandbox with a cap of two triggers per workflow and 20,000 a month on Professional with unlimited triggers per workflow. Sourcedocs.dify.ai node reference, dify.ai/workflows and dify.ai/pricingread 2026-08-31 |
|
| Model Flexibility & Routing | Full |
|
Model Providers is a documented configuration surface where the customer either spends Dify credits on hosted models or connects its own provider accounts and API keys, and the LLM node selects the model per node so different steps of one workflow can run on different models. The pricing page names OpenAI, Anthropic, Gemini, xAI and Tongyi as credit-metered options and states that once credits are exhausted the customer switches to its own key, model providers install as marketplace plugins, and air-gapped deployments connect locally deployed models. Sourcedocs.dify.ai model-providers, dify.ai/pricing and dify.ai/dify-enterpriseread 2026-08-31 |
|
| APIs, SDKs & MCP Extensibility | Full |
|
A REST API with a published OpenAPI specification integrates Dify applications into the customer's own product with app credentials, streamed responses and session management; difyctl is a documented CLI for running Dify apps from terminals, scripts, CI and other AI agents; a plugin development kit lets third parties build tools, models and integrations; published workflows are exposed as MCP servers consumable by Claude Desktop, Cursor and other MCP clients; inbound webhooks, custom endpoints and an External Knowledge API specification complete the surface, and the whole platform is open source on GitHub. Sourcedocs.dify.ai API reference and CLI overview, dify.ai/workflowsread 2026-08-31 |
|
| Testing, Debugging & Optimization | Partial |
|
A documented debug surface lets the customer run a single node or the whole flow against sample input, inspect intermediate variable values before real traffic arrives, review run history, read typed error metadata and configure predefined error-handling logic so nodes stop, return defaults or route to a recovery branch. Knowledge retrieval has its own hit-testing tool, an Annotation System curates approved answers with a per-tier quota, and version control restores any published build. No evaluation, scoring, regression-suite or benchmarking product appears anywhere in the 265-page cloud documentation index. Sourcedocs.dify.ai debug tree, knowledge test-retrieval and version-control referencesread 2026-08-31 |
|
| Browser & Computer Use | Not documented |
|
The complete node inventory in the cloud documentation index contains no browser, desktop or computer-use node. The nearest capabilities are an Import Data from Website crawl that feeds the knowledge base and a tutorial that scrapes a public profile through a third-party crawling tool, both of which reach content over HTTP rather than operating a rendered interface. Dify acts on external systems through APIs, HTTP requests, MCP tools and installed plugins, all of which are programmatic interfaces. Sourcedocs.dify.ai node reference and knowledge import treeread 2026-08-31 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Recent platform changes
Dify launched New Agent, which allows agents to operate as standalone and independent applications. These agents can also be utilized as reusable resources within larger workflows.
Bears on: Agent capability
View sourceDify introduced new agent features including an E2B Sandbox for secure code execution, Home Snapshots, and Skill Management. These updates expand the execution environment and tool orchestration capabilities for agents.
Bears on: Agent capability
View sourceDify released version 1.16.1, adding a Tool Multi-Select Input for configuring multiple tool parameters simultaneously and a Workflow Node Locator that links run-log errors directly to the corresponding canvas node. The update also shifts the default OpenAI plugin API from Chat Completions to Responses to support the newly released GPT-5.6 model family.
Bears on: Workflow orchestration
View sourcePricing
From $59/mo · Sandbox free demo + free self-host
credits
Included quota
Sandbox (free): 200 message credits stated without a monthly qualifier, 1 workspace, 1 member, 5 apps, 50 knowledge documents, 50MB knowledge storage, 10 knowledge requests/min, standard document processing, 3,000 trigger events, up to 2 triggers per workflow, standard workflow execution, 10 annotations, 30 days log history, 5,000 API calls/month. Professional ($590/workspace/yr): 5,000 message credits/month, 1 workspace, 3 members, 50 apps, 500 documents, 5GB storage, 100 knowledge requests/min, priority document processing, 20,000 trigger events/month, unlimited triggers per workflow, faster workflow execution, 2,000 annotations, unlimited log history, no Dify API rate limit. Team ($1,590/workspace/yr): 10,000 credits/month, 1 workspace, 50 members, 200 apps, 1,000 documents, 20GB storage, 1,000 knowledge requests/min, top priority document processing, unlimited trigger events, unlimited triggers per workflow, priority workflow execution, 5,000 annotations, unlimited log history, no API rate limit. Enterprise (custom): scalable deployment, commercial license authorization, multiple workspaces and enterprise management, SSO, negotiated SLAs, advanced security and controls, official updates and maintenance, professional support. Community: free, all core features in the public repository, single workspace, subject to the Dify open-source license.
What is public
Dify, by LangGenius, publishes four cloud tiers metered by message credits, plus a free self-hosted Community Edition. Sandbox is free with 200 message credits, 1 member, 5 apps and 50 knowledge documents. Professional is $590 a year per workspace (about $59 a month) with 5,000 credits a month, 3 members and 50 apps. Team is $1,590 a year (about $159 a month) with 10,000 credits a month, 50 members and 200 apps. Enterprise is custom and carries SSO and the enterprise security controls. Customers can bring their own model provider keys and pay the provider directly.
Billing mechanics
Fixed cloud tiers with included message credits. A credit is one model call on a Dify-hosted model and consumption varies by model type; connecting your own API key bypasses the credit system, so the subscription then covers only the platform. Tiers also differ by members, apps, knowledge storage and documents, and rate limits, and annual billing saves about 17%.
Cost watchouts
Sandbox's 200 message credits carry no monthly qualifier and read as one-time, so it is a demo rather than a free production tier; message credits meter MODEL CALLS not user messages and are consumed at different rates by model type, so budget several times a naive estimate; LLM inference is NEVER included, since you either spend credits or pay your provider directly under BYOK, and inference is often the largest line; SSO IS ENTERPRISE-ONLY as of the 2026-08-31 page, not Team as previously recorded, which moves single-sign-on from a $1,590/yr line item to a sales conversation; trigger events are now a metered quota, so scheduled and webhook-driven automation consumes plan capacity independently of message credits; the modified Apache-2.0 license BANS multi-tenant resale, so embedding self-hosted Dify in a SaaS you sell requires a commercial license from LangGenius.
Variable cost rationale
The Dify platform fee is fixed/predictable per tier, but real cost is driven by LLM provider usage - either via Dify credits (model-call-metered, model-choice-sensitive) or BYOK (you pay the provider directly); credits give predictability while BYOK shifts cost off Dify entirely
Additional watchouts
Read the license: under the modified Apache 2.0 terms, running self-hosted Dify as a customer-facing multi-tenant SaaS needs a commercial license from LangGenius. Sandbox's 200 credits carry no monthly qualifier, so treat it as a trial rather than a production tier. Credit use varies by model, self-hosting is free but carries Docker, operations and monitoring overhead, and model costs paid to providers can exceed the platform fee.
Overage / add-ons
When monthly credits run out, you upgrade a tier OR configure your own LLM API key (BYOK) to bypass credits entirely (paying the provider directly); Professional resources (vector storage, seats) can be topped up independently. Self-hosted Community Edition has NO credit system at all. Enterprise is custom.
Sales call required
Mixed (some tiers require a call)
Free / trial
Free
Lowest paid plan
$59/mo or $49.17/mo AE
Commercial notes
Dify is open source on GitHub with a large community (about 139K GitHub stars by recent counts) and more than 1M deployed apps, combining a visual workflow builder, RAG, an agent framework, model management and observability as a backend-as-a-service. The Community Edition uses a modified Apache 2.0 license that restricts multi-tenant commercial self-hosting.
Key ambiguities
SSO is listed under Enterprise only. Sandbox's 200 message credits are stated with no monthly qualifier while paid tiers read per month, which suggests a one-time allowance but is not stated outright. Enterprise has no public figure; an AWS Marketplace listing has been cited at about $150K a year, unverified. Credit rates per action beyond the statement that consumption varies by model type are not published, and the $59 and $159 monthly figures are inferred from the stated 17% annual saving.
Cancellation / refund
Sandbox free (no card, one-time credits); Professional/Team self-serve monthly or annual (annual ~17% cheaper; corporate payment methods supported); top up vector space/seats on Professional independently; Enterprise custom-contracted; self-hosted Community Edition is free (open-source, no contract - but multi-tenant commercial use needs a LangGenius license); students/educators free
Support SLA / resale
Community support (Sandbox and self-host); priority processing and execution tiers on Professional and Team; Enterprise adds negotiated SLAs by Dify partners, official updates and maintenance, and professional technical support, with the enterprise page naming a dedicated CSM, a Slack-connect channel, one-hour Sev-1 response and quarterly business reviews. SSO, RBAC, SCIM, session policies, tamper-evident audit logs and multi-tenant isolation are Enterprise capabilities. Connects to OpenAI, Anthropic, Gemini, xAI and Tongyi on credits, or bring your own provider keys, or locally deployed models in air-gapped installs; publishes MCP servers. Self-hosted multi-tenant resale requires a commercial license from LangGenius.
Missing data
Whether Sandbox credits renew monthly is not stated; Enterprise pricing is custom, with an unverified AWS Marketplace figure of about $150K a year; credit rates per action by model are not published.
Related vendors
- Lindy — AI teammate that lives in Slack, answers from your tools and…
- Activepieces — Open source, MIT licensed automation platform where every one of 700…
- AgentX — No code platform to build, evaluate, and deploy multi agent…
- AI Library — No-code platform for building and running AI agents, with a…
- Aigensei — Enterprise agentic workflow platform that runs business processes…
- Airia — Enterprise platform that unifies agent orchestration, security, and…
Alternatives to Dify
The closest documented capability profiles to Dify among agent builders tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- Activepieces11.5 / 14A lighter documented profile than Dify
- OutSystems12.5 / 14Fuller documented coverage on Testing, Debugging & Optimization
- AgentX12.0 / 14Fuller documented coverage on Testing, Debugging & Optimization
- Airia12.0 / 14Fuller documented coverage on Testing, Debugging & Optimization
- Airtable11.0 / 14A lighter documented profile than Dify
- Cohere North11.0 / 14A lighter documented profile than Dify
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded