Dify

What it is

Dify is a platform for building LLM applications: chatbots, retrieval (RAG) assistants over your own documents, multi-step AI workflows and, since July 2026, autonomous agents. You build on a visual canvas, plug in any model provider (OpenAI, Anthropic, Gemini, open-weight models through Ollama or vLLM), and publish the result as a web app, an embeddable chat widget, a REST API or an MCP server. Knowledge bases, prompt versions, run logs and annotation-based answer correction all live in the same workspace.

It is built by LangGenius, founded in 2023 and headquartered in Menlo Park. On March 9, 2026 the company announced a $30M Series Pre-A led by HSG at a $180M valuation, and said more than 280 enterprises run its commercial editions, naming Maersk and Novartis among them. The GitHub repository has more than 158,000 stars. Releases ship roughly every two to four weeks. The current version, 1.17.1, landed on September 10, 2026, after 1.16 put the new Agent App into open beta with a built-in sandbox and a skill system.

The license is “open source” with two conditions on top of Apache 2.0. You may not run the code as a multi-tenant service (one tenant is one workspace) without written permission, and you may not remove the Dify logo or copyright from the console or web apps. Internal use, including as the backend for your own product, is allowed.

Why it shows up in ops stacks

• RAG is the core, not an add-on. You upload contracts, policies, playbooks or job-description libraries, choose a chunking strategy and embedding model, and get a knowledge base that any app or workflow node can query, with hit testing to see which chunks came back. For a Legal Ops team building a policy Q&A bot, or a RevOps team grounding a deal-desk assistant in pricing rules, this is most of the build.
• MCP in both directions. Any Dify app can be published as an MCP server from its Access Point tab, so Claude Desktop or Cursor can call it as a tool. Dify can also consume outside MCP servers over HTTP transport and hand their tools to agents and workflow nodes. Version 1.16 added support for the 2025-06-18 MCP spec.
• Model choice per node. One workflow can route classification to a cheap model and drafting to a frontier model, which matters when the bill is per token.

Pricing reality

From dify.ai/pricing, checked 2026-10-11. Cloud plans are priced per workspace, not per seat.

• Sandbox, free. 200 message credits in total (not per month), 1 member, 5 apps, 50 knowledge documents, 30 days of logs.
• Professional, $59/month or $590/year (about $49/month). 5,000 message credits a month, 3 members, 50 apps, 500 documents, 5GB of knowledge storage.
• Team, $159/month or $1,590/year. 10,000 credits a month, 50 members, 200 apps, 1,000 documents, 20GB.
• Community Edition, free, self-hosted. All core features, one workspace. The documented minimum is 2 CPU cores and 4 GiB of RAM.
• Enterprise, custom. Multiple workspaces, SSO, a commercial license and support. No public figure, and we found no reliable third-party price data, so treat it as a sales conversation.

Message credits only pay for models that Dify hosts, and they burn fast on frontier models: the pricing page lists 20 credits per response on Claude Opus 4.7 and 5 on GPT-5, against 1 on Gemini 2.5 Flash. That makes Professional’s 5,000 credits about 250 Opus responses. You cannot buy more credits. Once they run out you switch to your own API key and pay the provider directly. In practice, every serious deployment ends up as $590–1,590 a year in platform fees plus a separate model bill. The Team plan is the realistic ceiling for one department: 50 members for $1,590 a year is far cheaper than any per-seat agent platform.

Best for

• Best for: an ops engineer or technical analyst at a 50–1,000 person company who needs to ship two or three internal AI apps, such as a policy Q&A bot, a contract-clause extractor or a candidate-screening assistant, grounded in the team’s own documents and owned by the team rather than IT.
• Legal Ops and Recruiting teams that must keep documents on their own infrastructure: the Community Edition runs on a single small VM.
• Not for teams whose real job is moving records between SaaS apps on a schedule. Dify has triggers and HTTP nodes, but its connector catalogue is thin next to a dedicated automation engine.

Versus the alternatives

• n8n: the share leader for technical automation, with 1,000+ integrations and per-execution billing from €20/month. Pick n8n when the work is CRM-to-enrichment-to-Slack plumbing with an AI step in the middle. Pick Dify when the work is the AI app itself: retrieval quality, prompt iteration and a chat front end. Many teams run both, with n8n calling Dify’s API.
• Microsoft Copilot Studio: the other share leader, $200 a month per 25,000-credit pack. Pick it when the company lives in Microsoft 365 and IT wants agents inside that tenant’s governance.
• Gumloop: the fastest-growing AI-native builder, Pro at $37/month with unlimited seats. Pick it when nobody on the team will touch a model provider setting. Pick Dify when you need self-hosting or control over chunking and retrieval.
• Sim: the younger open-source rival, Pro at $25 per user per month. Pick it if you prefer a fully open, agent-first canvas and can live with a smaller ecosystem.
• Relevance AI: pick it when the goal is a team of sales agents working inside your GTM tools, not a custom app.
• Dust or Langdock: pick these when what you want is a company-wide chat seat with agents, rather than apps a builder publishes to others.
• Flowise (now owned by Workday) and Langflow (now IBM, via DataStax) are the other open-source canvases. Both are now inside larger vendors, which is worth weighing if independence matters.

Watch-outs

• The license is not plain Apache 2.0. Running Dify as a multi-tenant service, or white-labelling the front end, needs a commercial license. Guard: if you plan to give each client its own workspace, or to ship Dify screens under your brand, get written terms from LangGenius before you build. If you only need the API behind your own UI, the frontend clause does not apply.
• Credits run out quickly on frontier models. At 20 credits per Opus response, a busy internal bot can exhaust Professional in days. Guard: add your own provider keys on day one and set them first in each provider’s Usage Priority, so credits are only a fallback and model spend appears on an invoice you control.
• Upgrades break self-hosted installs. The 1.17.1 notes require a staged upgrade for the bundled Weaviate store, and 1.16.1 added a mandatory new environment variable. Guard: pin a version, read the release notes before each upgrade, and test on a staging copy of your data volumes before touching production.
• Agent App is still in beta. Guard: keep production use cases on Workflow and Chatflow apps, which are the mature app types, until Agent leaves beta.