Dify

Dify: an open-source platform you can buy, host or both

Dify ranks second on this site, and for a platform leader it is the contender that most literally offers build and buy in one product. Teams can run the source themselves, pay for the hosted cloud, or license an enterprise edition that runs inside their own perimeter. Dify describes itself three ways depending on where you look: its docs speak of an open-source platform for AI applications¹, its repository of an open-source environment for developing LLM apps², and its home page of a platform built for agentic workflows that are ready for production³.

What follows reviews Dify against the question this site asks: how much of the agent work does a team still have to do itself, and what does it get for doing it? The short answer is that Dify gives an engineering team more control over hosting than any other contender here, and asks that team to do the build.

Scale and license

The home page lists the self-hosted Community Edition with a count of 148K+ stars⁴. On GitHub the counter read 160.7k on our check date⁵, a higher figure than the home page's, and neither page accounts for the difference. Either figure is well above the star counts shown for LangGraph and CrewAI on their repositories^6,7.

The license deserves a lawyer's attention. The file itself starts from Apache 2.0 and then adds conditions of its own⁸; the marketing site labels the result an Apache-2.0-derivative license⁹. The condition that matters most to a platform team is the one forbidding a multi-tenant service built on the source unless Dify grants written permission¹⁰. A team that plans to offer Dify to several business units as separate tenants, or to resell it, should review that clause with counsel before committing.

How a build works

At the center of the product is Workflow Studio, a drag-and-connect canvas on which agentic workflows are laid out¹¹. A builder can talk an agent into existence through chat, or set every field manually if that suits the team better¹². The Knowledge Pipeline turns files, websites, online documents and drives into repeatable ingestion pipelines¹³. Engineers who like to tune retrieval will appreciate the classic knowledge base mode, in which chunking, indexing and search behavior are all left to the team's choice¹⁴.

The GitHub page promises autonomous agents an isolated workspace of their own for commands and file handling¹⁵. Application logs and performance history are kept for analysis¹⁶, and the README states that each part of the product can be driven through an API¹⁷. When an app is finished, Dify can expose it four ways: as a hosted app, an API endpoint, an embed in another page, or a tool that speaks MCP¹⁸.

That last point is important for the build-versus-buy decision. A team can build a workflow visually in Dify and then call it from its own code as an API or an MCP tool. Dify can serve as the agent layer inside a larger system that engineers still own.

Running it yourself

The self-hosting documentation is specific, which is a strength in itself. The self-host overview sets a low bar for where Dify will run, naming laptops, on-premises servers and cloud VMs, in effect any machine with Docker¹⁹. Hardware needs are modest: the guide's minimum is a two-core machine with 4 GiB^20,21, running Docker Compose at version 2.35.0 or later²². Mac users running Docker Desktop are told to allot the VM at least two virtual CPUs and double that memory, 8 GiB²³. For the Community Edition, the site promotes a Docker install that takes one command²⁴, and teams on Kubernetes get a Helm chart²⁵.

Out of the box the Compose stack ships its own PostgreSQL and Weaviate along with supporting services²⁶, and keeps the database, Redis, sandbox and vector store off host ports by default, which a security reviewer will welcome²⁷. One step is easy to miss on a first install: before the first launch, generate the shared secret the guide asks for, because the agent backend and the local sandbox refuse to boot without it²⁸. Once the containers are up, an operator opens the install page on localhost and creates the admin login there²⁹. Two operator warnings stand out in the guide: swap out the development credentials that ship in the environment file³⁰, and apply upgrades with up -d, since a plain restart carries the previous container environment forward³¹.

Two operating limits deserve attention from whoever will run the stack. The docs are blunt about it: alter DB_TYPE or VECTOR_STORE and nothing is migrated³², so a switch made after the first launch strands the data already written³³. A platform team should choose its database and vector store before the first production deployment and treat that choice as fixed.

Integrations

Integration in Dify is organized as plugins, which can bring in a model provider, a tool, a data source or an outside service³⁴. Teams can pull ready-made ones from the Marketplace, MCP integrations included, or write their own definitions³⁵. What an agent can call is broad, ranging from tool plugins and MCP servers to other published workflows and anything described by OpenAPI³⁶. Quality control on the Marketplace is split: Dify reviews and tests each listing before it goes up, and the publisher maintains it afterward³⁷.

For engineering teams, the install paths are flexible and reach beyond the store: a public GitHub URL plus a version tag will install a plugin³⁸, and internal ones can be shipped as a local .difypkg package³⁹. If you run your own moderation service or data API, a custom endpoint wires it into Dify apps⁴⁰. For knowledge ingestion, the examples given are content held in Google Drive and in Notion⁴¹.

Dify does not publish a total connector count on the pages opened for this review. Two of its rivals here do publish totals, with Nexus above 4,000⁴² and Relevance AI above 2,000⁴³. A Dify buyer should expect to write or install some connectors, which is consistent with a platform aimed at teams who build.

Security and certifications

Dify Enterprise carries two named attestations on the home page, SOC 2 Type II and ISO 27001⁴⁴. The enterprise page fills in the controls a security architect will ask about: single sign-on over SAML or OIDC⁴⁵, audit logs that are tamper-evident and flow into your own SIEM⁴⁶, and encryption under keys you supply⁴⁷. In its privacy policy, Dify sorts its safeguards into technical ones, such as encrypting traffic and stored data, controlling access and watching for intrusions⁴⁸, and human ones, chiefly training staff to protect personal data⁴⁹. Security researchers are directed to file findings privately via GitHub's advisory system⁵⁰.

Those are among the most complete security statements on this site. They apply to the Enterprise edition; a team running the Community Edition takes on its own hardening, starting with the credential replacement the guide calls for³⁰.

Governance and permissions

Workspace roles are documented plainly. Each workspace has a single Owner, who holds every permission⁵¹. Below the Owner, an Admin can do all that an Editor does and can also handle membership and model provider settings⁵². Editors have full rights over apps and knowledge bases, creation and deletion included⁵³, while the Normal role is limited to using whatever has been published⁵⁴. Housekeeping rules are sensible: an invite is valid for 72 hours⁵⁵, removal cuts access immediately⁵⁶, and handing over ownership requires a verification code sent by email to the current Owner⁵⁷. Control over plugins rests with the Owner and Admins, who choose which members may install or debug them⁵⁸, with the debugging key gated by a permission of its own⁵⁹.

The Enterprise license extends that model to RBAC at the level of a single workflow, user provisioning over SCIM and session rules⁶⁰, and it lifts the one-workspace ceiling with central management across several⁶¹. The Community Edition is limited to a single workspace⁶².

What Dify does not describe on the pages opened is a release gate of the kind Nexus applies, where an agent cannot ship until a CIO signs off on data flow, access and audit⁶³. Dify supplies the controls; the team decides how to use them.

Where your data lives

Deployment choice is Dify's clearest advantage for a platform team, and the enterprise page lists three self-managed targets: your VPC, your data center, or a network cut off from the internet⁶⁴. Teams wanting dedicated infrastructure that someone else runs can pick the Dify-managed option, a single-tenant cluster placed in the region they name⁶⁵. The self-managed VPC route lands in your own account on AWS, GCP or Azure, and Dify's claim for it is that data stays within your perimeter throughout⁶⁶. In the air-gapped variant, nothing needs an outside connection, because models can be hosted locally alongside the platform⁶⁷. Its residency promise, in short, is that your VPC, your region and your perimeter bound where data goes⁶⁸.

Dify Cloud is the hosted alternative, with no infrastructure to manage⁶⁹. On training, the privacy policy makes two commitments: Dify trains no AI models itself, and the policy rules out using what customers send through the product as training material⁷⁰. Data flow depends on who picks the model: a customer-chosen model with the customer's key sends interaction data directly to the provider for processing⁷¹, whereas a Dify-chosen model puts that provider in the role of Dify's sub-processor⁷². A cloud buyer with residency obligations should note one broad clause, which allows storage and processing wherever LangGenius, an affiliate or a service provider runs facilities⁷³, and should weigh the enterprise deployments against it.

Pricing

Dify publishes prices, which makes budgeting a pilot simple. The Professional plan costs 590 dollars per workspace per year⁷⁴ and the Team plan 1,590 dollars⁷⁵. Paying yearly instead of monthly cuts the cloud bill by 17 percent⁷⁶. Message credits come in different sizes by tier, a one-off 200 on the free Sandbox⁷⁷ against a monthly 5,000 on Professional⁷⁸. Once the allowance is spent, a team can plug in its own model API key and continue⁷⁹. The Enterprise column on the pricing page simply reads Custom⁸⁰.

Trigger events are rationed on the lower tiers, with Sandbox allowed 3,000⁸¹ and Professional 20,000 per month⁸², and the cap disappears on Team⁸³. Upgrades take effect immediately for the difference in price⁸⁴; downgrades start at the next billing cycle⁸⁵. Verified students and school staff can run the annual Professional tier at no cost⁸⁶, minus ticketed and priority email support⁸⁷.

The license fee is only part of a Dify budget. Self-hosting means the customer pays for its own infrastructure and the engineers who run it, and the build work itself sits with the customer. That trade is the main reason Dify sits below Nexus here, where the price is set per role automated and includes the build⁸⁸.

Plan limits

The lower tiers are sized for individuals and small teams. Sandbox is a single-user tier⁸⁹, throttled to ten knowledge queries per minute⁹⁰, with logs kept for thirty days⁹¹. Professional allows three members⁹². Knowledge storage is metered against the plan, and at the limit, classic and agent knowledge bases alike stop accepting content⁹³. Dify may reject a file up front if its estimated footprint is too big⁹⁴, and at the cap the only ways forward are deleting content or moving up a plan⁹⁵, though search and deletion keep working meanwhile⁹⁶. One migration path is closed: knowledge bases that are external or were built through Knowledge Pipeline stay as they are and cannot become agent knowledge bases⁹⁷.

Intended buyers

Dify's own segmentation is clear. It pitches Professional at individual builders and small groups putting AI apps into production⁹⁸. The Team tier targets mid-sized groups whose needs are collaboration and more throughput⁹⁹. Enterprise is the tier for organizations whose checklist includes advanced security, compliance and governance features plus a dedicated support line¹⁰⁰. As for the free self-hosted edition, Dify frames its audience as hobbyists of open source, single developers and projects with no commercial aim¹⁰¹.

Four sectors get solution pages. Universities are offered a single knowledge base serving several campuses¹⁰², public bodies a local install they can share securely¹⁰³, life-science firms governed assistants at expert level¹⁰⁴, and consumer-goods companies a way for business teams to make their own agents¹⁰⁵. The enterprise page carries a pair of customer metrics worth reading closely: one compares build time per app with the four months it took when the customer built alone¹⁰⁶, and the other counts business users who learned to build agents during FY2025¹⁰⁷. Both are customer claims published by Dify.

Build, buy or both with Dify

Dify offers all three on one codebase. To build, a team installs the Community Edition with one Docker command²⁴, accepting a single workspace⁶² and an edition Dify aims at hobby and non-commercial use¹⁰¹. To buy, a team signs up for Dify Cloud with no installation¹⁰⁸ or commissions a single-tenant cluster managed by Dify⁶⁵. To do both, it licenses Enterprise and runs it in its own VPC⁶⁶, with Dify supplying the product and the customer supplying the infrastructure.

Model costs stay under the customer's control in each case. Message credits can be replaced by the customer's own API key once used up⁷⁹, and with the customer's key, interaction data goes straight to the chosen provider⁷¹.

The build effort itself is the constant. In each mode the customer's team designs the workflows, connects the systems and owns the result. That is why a team that wants Dify's openness and a faster first agent might pair it with outside help, and why Nexus, which includes the build in its price per role automated⁸⁸, ranks above it here.

Verdict

For an engineering leader who wants to own the agent stack, Dify is the strongest pure platform on this site. Prices are public, the self-host guide is thorough, the Enterprise edition can sit in a VPC, a data center or an offline network⁶⁴, and that edition holds SOC 2 Type II and ISO 27001⁴⁴. It ranks below Nexus because the build, the operations and the connectors remain the customer's work, and because the license limits multi-tenant use¹⁰. Teams with spare engineers and a strict hosting rule should put Dify on the shortlist next to Nexus and set the full cost of the build beside the license fee.