Tool

Dify

Dify review: compare 2026 Cloud pricing, Community self-hosting, license restrictions, privacy, human approval, recovery, and governance.

Updated 2026-10-14

DifyAI applicationsworkflow automationself-hostedRAGAI agentshuman approval

VendorDify

Official siteVisit site

PricingView pricing

PlatformsWeb, API, Docker, Self-hosted

Free tierSandbox Cloud is free with 200 message credits; Community Edition can be self-hosted free under the Dify Open Source License

Starting price$59 per workspace/month for Professional; $590 per workspace/year with annual billing

Dify workflow canvas with knowledge retrieval, LLM, human approval, regeneration, forwarding, timeout, and answer branches — Dify's March 2026 first-party example routes a model-generated sales analysis through confirm, regenerate, forward, and timeout branches; this is vendor media, not our workflow test.

Verdict

Choose Dify when the thing being shipped is an LLM application: a chatbot, RAG service, agent, or API whose prompts, model calls, knowledge retrieval, tools, branches, and logs should live in one visual platform. It is a weaker default for ordinary SaaS-to-SaaS automation, where n8n has a broader integration-first operating model. Self-host Dify when model and application control justify owning a multi-service stack. Do not choose Community Edition on “open source” shorthand alone: its modified Apache 2.0 license restricts multi-tenant use and frontend-brand removal, while Enterprise holds the stronger multi-workspace, SSO, security, and support package.

Dify combines a workflow canvas, model-provider management, RAG pipelines, agents, monitoring, web apps, and service APIs. That makes it closer to an AI application runtime than a generic trigger-action connector.

Start with the best self-hosted AI automation tools when deployment control is the deciding factor, or the broader AI workflow automation shortlist when category fit is still open. Compare the n8n profile for integration-heavy business processes. If an existing stack includes Flowise, use the Flowise profile for its end-of-life and migration context. The reliable no-code AI workflow guide covers schemas, evaluation, approval, idempotency, and monitoring across platforms.

Decide whether Dify is the right kind of workflow system

Criterion	Dify fits when	Validate before rollout
Product shape	The output is a model-backed app or API with prompts, retrieval, tools, or agent behavior	Build one real request path, including bad input, empty retrieval, provider timeout, and malformed model output
Integration depth	HTTP, plugins, MCP, and supported tools cover the required systems	Inventory every source and write target; compare gaps and credential ownership with n8n
Knowledge	Documents and retrieval are part of the application, not a separate data service	Test chunking, permissions, deletion, re-indexing, citation quality, and recovery with representative documents
Billing	Workspace, app, storage, trigger, request, and model costs remain predictable	Count model responses, trigger events, knowledge operations, external-provider charges, and retained data separately
Deployment	The team can operate the API, workers, database, Redis, file storage, vector store, plugin services, and upgrades	Restore a backup, rotate secrets, replay a failed run, and stage the exact vector-store upgrade path
Governance	One Community workspace is sufficient, or Enterprise controls are in the signed package	Map builders, publishers, model-key owners, reviewers, log readers, and workspace administrators to entitlements
Human review	Web-app or email approval can pause a workflow before a sensitive action	Test authentication, leaked links, first-response races, timeout, duplicate submission, upgrade, and resume behavior

Cloud price is only one part of cost

The Dify Cloud pricing page was checked October 14, 2026. Monthly and annual options differ; the annual prices below are paid per workspace for the year.

Plan	Public price	Included model allowance	Operational limits that change the decision
Sandbox	Free	200 message credits	1 member, 5 apps, 50 knowledge documents, 50 MB knowledge storage, 10 knowledge requests/minute, 3,000 trigger events, 2 triggers/workflow, 30-day logs, 5,000 API calls/month
Professional	$59/month or $590/year	5,000 message credits/month	3 members, 50 apps, 500 documents, 5 GB knowledge storage, 100 knowledge requests/minute, 20,000 trigger events/month, unlimited log history, no Dify API rate limit
Team	$159/month or $1,590/year	10,000 message credits/month	50 members, 200 apps, 1,000 documents, 20 GB knowledge storage, 1,000 knowledge requests/minute, unlimited trigger events, unlimited log history, no Dify API rate limit
Enterprise	Custom	Contract-specific	Self-hosted enterprise package with multiple workspaces, SSO, advanced controls, support, and commercial-license authorization

Message credits are starter model access, not a fixed number of tokens or complete application runs. Dify says one model response consumes a model-dependent number of credits; a workflow with several model nodes can consume several responses. When credits run out, a workspace can use its own provider key. The provider then bills the account directly, and tool APIs, storage, vector infrastructure, observability, and self-host operations remain separate costs.

The Sandbox page publishes 200 credits and 3,000 trigger events without stating a reset period beside either item. Do not model them as monthly allowances unless the checkout or workspace meter confirms that. Professional and Team explicitly label their message credits per month; Professional explicitly labels trigger events per month.

Community is free to run, with license and governance boundaries

Dify Community Edition is free to self-host and exposes the core public-repository features in one workspace. Enterprise adds multiple workspaces and management, SSO, advanced security and controls, scalable deployment options, official maintenance, and support. Confirm every required control in the order form; “self-hosted” describes location, not role design, audit coverage, high availability, or vendor responsibility.

The repository is licensed under the Dify Open Source License, a modified Apache 2.0 license. Its text allows commercial use, including use as a backend or enterprise application-development platform, but requires a commercial license for an unauthorized multi-tenant environment. It also prohibits removing or modifying the Dify logo or copyright information when using Dify's frontend. The pricing page describes Community as suited to non-commercial projects, while the license text permits broader commercial use subject to those restrictions. Review the actual architecture and license text rather than treating either marketing copy or “Apache-based” as the whole answer.

Self-hosting also creates an upgrade and recovery job. Dify's Docker deployment includes application services plus database, Redis, storage, vector-store, sandbox, proxy, and plugin components. Back up the database, uploaded files/object storage, vector data, configuration, and secrets; then test a restore at the pinned application version. Read release-specific migration notes before changing images. The v1.17.1 release warned that bundled Weaviate users needed a manual staged upgrade across 12 minor versions and that a pull-and-restart path could silently and permanently break vector search. That warning is version-specific, but the operating lesson is durable.

Privacy depends on deployment and every connected provider

Dify's current privacy policy says Cloud collects prompts, uploaded files, knowledge-base content, outputs, application configurations, and workflows. LangGenius says it does not use AI interaction data to train models. When the customer supplies a model-provider key, interaction data goes directly to that provider and its agreement controls model training and retention. When LangGenius manages the provider, it says contracted subprocessors cannot use the data for training or their own purposes.

For on-premise deployments, Dify says it does not collect ordinary product-usage information through the product, though license verification and update services may send basic connection and version data. Self-hosting still does not imply zero egress: configured model providers, tools, plugins, email delivery, external knowledge sources, and observability services receive the data the workflow sends. Build a node-by-node data-flow inventory and test outbound network policy rather than relying on the deployment label.

Human Input is useful, but the approval link is a security boundary

The Human Input node pauses a workflow, displays dynamic content and form fields, and routes the next step through named actions. It can deliver through the current web app or email. Web-app delivery is unavailable for workflows started by a Trigger; email can target workspace members or external addresses. Current documentation says anyone with the emailed link can respond and the request closes after the first response. Treat that URL as a credential, show the exact proposed action, and recheck mutable state after approval.

The default timeout is three days. Builders can connect a timeout branch to a reminder, escalation, or safe termination; without one, the workflow ends. Test approve, reject, edit, forward, timeout, duplicate response, first-response race, expired link, and worker restart. A review node does not make downstream writes idempotent.

Dify also offers node-level recovery for LLM, HTTP, Code, and Tool nodes: stop the workflow, substitute a typed default value, or follow a fail branch. Loop nodes stop on a child failure; Iteration nodes can terminate, continue with null, or remove failed outputs. Those controls are useful, but a default value can conceal a material failure and removing abnormal outputs changes result cardinality. Send error_type and error_message to observable failure handling, and keep external writes deduplicated.

One resolved HITL report shows why version acceptance matters

In a GitHub issue opened September 9, 2026, iamyz reported that Dify v1.17.0 on self-hosted Docker accepted Human Input in the design window but left the published web window unable to accept further input after the button click. The issue was assigned to the 1.17.1 milestone and closed through PR #42082. Dify's v1.17.1 release notes say the fix restored the completion events needed to clear the approval UI and unblock published Web App input.

That is one reproducible, resolved, version-specific report. It is not evidence that current HITL is generally broken or reliable. It does justify pinning a release and running the exact published-app approval path after every upgrade.

We did not create a Dify Cloud workspace, deploy Community Edition, run a workflow, submit an approval, restore a backup, inspect network traffic, audit source code, or benchmark latency and reliability for this profile. Pricing, limits, controls, and defects above come from current first-party pages and the attributed issue trail.

Choose Dify when one platform should own an LLM application's prompts, retrieval, model routing, tools, workflow, API, and runtime evidence. Choose n8n when broad application integration is the center of gravity. For an existing Flowise deployment, read the Flowise profile for end-of-life dates and migration planning before committing more work to that stack.