Dify

Dify: an open-source visual builder that can live inside your own network

Dify occupies the middle of the build spectrum, and it reaches toward both ends. The documentation frames Dify as open-source software for making AI applications¹. GitHub frames it as an open-source platform for developing LLM apps². The home page title goes furthest, promising agentic workflows ready for production³. In practice the product is a visual workflow studio that business users can learn, with APIs, plugins and a self-hosted edition that engineers can take apart. For a Head of Digital, Dify is the builder on this site that gives the most control over where the software runs, and that is the main reason it ranks third.

This page looks at Dify through the questions a buyer choosing a build style will ask: who can build on it, how it is installed and run, what governance a central team gets, and what the plan limits mean for a department that starts on the cloud and later wants to move in house.

The open-source position

Dify's community footprint is large. The home page lists the Community Edition as source-available with 148K+ stars and a Docker self-host route⁴, while the GitHub page showed 160.7k stars when checked⁵. The two figures differ because they were captured on different pages.

The license deserves a careful read before an enterprise relies on it. The code ships under Apache License 2.0 with extra conditions attached⁶, which the home page describes as a license derived from Apache-2.0⁷. A key condition: running the source as a multi-tenant service needs written permission first⁸. For most internal deployments that will not matter. For a company that wants to offer agents built on Dify to its own customers as a shared service, it is a point for legal review.

How building works

The core of the product is Workflow Studio, described as a visual builder for agentic workflows⁹. Makers can either talk an agent into shape in a chat with the builder or set every option by hand¹⁰, which gives a business user a conversational start and a technical user fine control. The GitHub README describes autonomous agents with a sandbox of their own that runs commands and handles files¹¹.

Knowledge is handled through two routes. With Knowledge Pipeline, a team sets up ingestion it can repeat for drives, files, websites and online documents¹². In a classic knowledge base, chunking, indexing and search settings are all the customer's choice¹³. That level of control appeals to technical teams who want to tune retrieval, and it asks more of a business user than a simple upload.

Output can take several forms. A finished app can go out as a hosted experience, an embed, an API endpoint or an MCP-compatible tool¹⁴, and each Dify offering has a matching API¹⁵. The README notes that application logs and performance can be monitored and analyzed over time¹⁶. For a build-style decision, the API coverage is what lets Dify serve both camps. A business team can build in the studio, and a development team can call the result from its own code.

Integrations and the plugin model

Integrations connect Dify to model providers, tools, data sources and external services¹⁷. The Marketplace installs model providers, tools, data sources and MCP integrations, or a team can define its own¹⁸. Agents can call tool plugins, OpenAPI services, MCP servers and workflows already published¹⁹. Knowledge sources include Google Drive and Notion pages²⁰.

The plugin model is unusually open. Marketplace integrations are reviewed and tested before publication, then maintained by their publisher²¹. Beyond the Marketplace, a team can pull an integration from a public GitHub repository by giving its URL and version²², or load an internal one packaged as a local .difypkg file²³. Custom endpoints connect a company's own API for moderation or external data tools²⁴. A Head of Digital with an engineering team will see this as flexibility. A security lead will want a rule on who may install from public repositories, and Dify provides one: the workspace owner and admins control who can install and debug integrations²⁵, with a separate debug permission covering the debugging key²⁶.

Dify's integration pages describe a model in which the customer adds what it needs^17,22. Nexus takes the catalogue approach and publishes a count of more than 4,000 integrations²⁷.

Triggers are metered by plan. The Sandbox plan includes 3,000 trigger events²⁸, Professional includes 20,000 a month²⁹, and Team lists unlimited trigger events³⁰.

Setup: cloud or your own servers

There are two ways to start. Dify Cloud is hosted by Dify, so teams can build, test and launch AI apps without managing infrastructure³¹. Signing up provides the managed service at once, with nothing to install and a free Sandbox tier³².

The self-hosted route is documented step by step. Dify runs anywhere Docker runs: a laptop, an on-premises server or a cloud VM³³. For the Community Edition, the home page offers Docker deployment in one command³⁴ and a Helm chart for Kubernetes³⁵. The Docker Compose guide asks for Docker Compose 2.35.0 or later³⁶, at least 2 CPU cores³⁷ and at least 4 GiB of RAM³⁸. Mac users are told to allot Docker Desktop no less than 2 virtual CPUs and 8 GiB of memory³⁹. The agent backend and its local sandbox require a shared token generated before the first start⁴⁰. Once the stack is running, the first admin account is set up through the install page in a browser⁴¹.

The guide includes operating advice that signals who the self-hosted edition is for. Upgrades should use up -d, because restart keeps the containers' old environment⁴². The default stack bundles PostgreSQL, Weaviate and the other services it needs⁴³ and publishes no host ports for the database, Redis, code sandbox or Weaviate⁴⁴. The bundled development credentials in docker/.env should be replaced⁴⁵. Data stays where it is if DB_TYPE or VECTOR_STORE is changed⁴⁶; a switch made after first start leaves existing records behind⁴⁷.

These are reasonable instructions for a platform team. They also show that self-hosting Dify is an engineering commitment: someone must run the containers, rotate credentials, manage upgrades and choose the database and vector store before the first start. Nexus handles delivery differently: its consultants build next to the customer, and production starts at week 4 of a 3-month proof of concept (owner decision⁴⁸).

Enterprise hosting options

Dify Enterprise offers a wide set of hosting choices. It can be deployed inside the customer's own VPC, on-premises or on an air-gapped network⁴⁹. A Dify-managed dedicated option runs on a single-tenant cluster in the customer's region⁵⁰. With the customer-managed VPC option, Dify runs inside the buyer's own AWS, GCP or Azure account, and the vendor says data stays within that perimeter⁵¹. Air-gapped installs work entirely offline, connected to models hosted on site⁵². Data residency is described as keeping data inside your VPC, your region and your perimeter⁵³.

For a company whose policy requires agent software inside its own cloud account, these options will often decide the shortlist on their own.

Security and certifications

For Dify Enterprise, the home page cites SOC 2 Type II together with ISO 27001⁵⁴. Enterprise identity uses SSO via SAML or OIDC⁵⁵. Audit logs are tamper-evident and streamed to the customer's SIEM⁵⁶, and encryption supports bring-your-own-key⁵⁷. The privacy policy lists encryption of communications and data, access control, and detection and prevention of unauthorized access as technical measures⁵⁸, and employee training on protecting personal information as a human measure⁵⁹. Vulnerabilities are reported privately through GitHub Security Advisories⁶⁰.

On data use, the privacy policy states that Dify trains no models of its own and keeps AI interaction data out of any training⁶¹. With a customer-chosen model provider and API key, interaction data goes directly to that provider⁶²; with Dify-selected models, the providers act as Dify's sub-processors⁶³. It also allows storage and processing wherever LangGenius, its affiliates or its service providers have facilities⁶⁴. A buyer using Dify Cloud should read that line together with the enterprise residency options.

The Dify home page names SOC 2 Type II and ISO 27001⁵⁴. Nexus holds SOC 2 Type 2, GDPR, ISO 27001 and ISO 42001 (owner decision⁴⁸).

Governance and roles

Workspace roles are simple and clearly defined. The Owner has full control of the workspace, and there is one per workspace⁶⁵. Admins manage members and model providers, plus everything an Editor can do⁶⁶. Editors create, edit and delete apps and knowledge bases⁶⁷. The Normal role uses published apps only⁶⁸. An invitation lapses after 72 hours⁶⁹; a removed member loses access at once⁷⁰; and handing over ownership needs a code sent to the Owner by email⁷¹.

Enterprise adds fine-grained RBAC down to workflow level, SCIM provisioning and session policies⁷², along with multiple workspaces and enterprise management⁷³. For a Head of Digital who plans to let several business units build, workflow-level permissions and SCIM are the controls that make a shared platform manageable.

Pricing

The cloud plans are priced per workspace per year. Professional costs $590⁷⁴ and Team costs $1590⁷⁵. Annual billing saves 17%⁷⁶. The Sandbox plan is free with 200 message credits⁷⁷, Professional includes 5,000 message credits a month⁷⁸, and once credits are used up a team can switch to its own API key⁷⁹. Enterprise is custom priced⁸⁰.

Upgrades take effect immediately, with the difference charged for the rest of the current billing period⁸¹; downgrades take effect at the start of the next cycle⁸². Dify for Education gives students, teachers and staff the Yearly Professional plan free after verification⁸³, without ticket or priority email support⁸⁴.

The plan descriptions show the intended buyers. Professional serves solo developers and small teams shipping AI apps to production⁸⁵. Team is for medium-sized teams needing collaboration and higher throughput⁸⁶. Enterprise is for enterprises requiring advanced security, compliance, governance and dedicated support⁸⁷. Community Edition is aimed at fans of open source, solo developers and projects with no commercial aim⁸⁸.

Plan limits

Several limits shape how far a team can go before upgrading. Sandbox allows 1 member⁸⁹ and Professional allows 3⁹⁰. Sandbox allows 10 knowledge requests per minute⁹¹ and keeps 30 days of log history⁹². Community Edition offers just one workspace⁹³.

Knowledge storage is capped per plan. Dify may refuse an upload based on its estimated storage⁹⁴. At the cap, no knowledge base of either type accepts new content⁹⁵; the block holds until usage drops back under the limit or the plan goes up⁹⁶, while search and deletion carry on as normal⁹⁷. Knowledge bases that are external, or that were made with Knowledge Pipeline, have no upgrade path to the agent type⁹⁸.

Who uses it

The solutions pages name several sectors. Education buyers get one knowledge base for multi-campus services⁹⁹. Public-sector buyers can deploy locally and share securely across institutions¹⁰⁰. Pharma and life sciences buyers get governed expert-level AI assistants¹⁰¹. Consumer-goods buyers are offered a way to let business teams build their own agents¹⁰². The enterprise page quotes customer claims about per-app build time, down from 4 months self-built¹⁰³, and business users trained to build AI agents in FY2025¹⁰⁴.

Strengths

Dify's main strength is control of hosting: VPC, on-premises, air-gapped and single-tenant options are all documented^49,50,52. Its visual studio serves business users, while its APIs and plugin model serve engineers^9,15,22. Enterprise governance reaches workflow-level RBAC with SCIM⁷², audit logs stream to a SIEM⁵⁶, and SOC 2 Type II and ISO 27001 are listed⁵⁴. The cloud price list is public^74,75.

Limits

Self-hosting requires a team that can run Docker or Kubernetes and manage data stores^36,46. The license restricts multi-tenant use without authorization⁸. Lower cloud plans are tight on members and storage^89,90,95. Hands-on delivery is described as dedicated support on the Enterprise plan⁸⁷, so a buyer who wants build help should ask what that support covers.

Verdict

Dify ranks third. For a Head of Digital whose first requirement is running the builder inside the company's own perimeter, it is the strongest option on this site. It ranks behind Nexus because Nexus takes on the delivery work that Dify leaves to the customer. Nexus consultants build with the team, production begins at week 4 of a 3-month proof of concept, the price follows the role automated, and the certifications include ISO 42001 (owner decision⁴⁸). It also ranks behind Relevance AI, whose embedded deployment team¹⁰⁵ and published uptime SLA¹⁰⁶ serve a business-led buyer more directly.