Open-source LLM app development platform — build AI agents and workflows visually.
Agents
Dify is an open-source LLM app development platform that lets you build AI agents, RAG applications, and multi-step workflows visually — no framework code required for most use cases, and a code escape hatch when you need it. It launched in 2023 as a bet that the fastest way to a working AI agent should not involve reading LangChain documentation for a week. That bet paid off. In 2026, Dify is one of the three or four platforms serious teams evaluate when they need to ship an AI application in weeks, not quarters.
Dify is what happens when a founder team asks: what if the workflow builder for LLMs was as good as n8n's is for automation? You open the canvas, drag an LLM node, drag a knowledge-base node, wire in a conditional branch, add a code node for the edge case the UI doesn't cover, hit publish. You have an API endpoint your app can call in about the time it takes to install Python and clone a starter template.
The distinctive positioning versus LangChain and CrewAI is that Dify is a platform, not a library. LangChain gives you a Python SDK and expects you to build the runtime yourself. CrewAI gives you an orchestrator for multi-agent conversations. Dify gives you a hosted (or self-hostable) runtime with a UI, an API, an admin console, RAG built in, model routing, prompt versioning, and a cost dashboard. The tradeoff is that you're committing to Dify's node model — but the payoff is time-to-production that no code-first framework matches.
We use Dify in production at AIQORA, which is why it's on this list. It has flaws — we'll get to them — but the productivity delta over hand-rolling a LangChain agent is genuinely large.
Visual Workflow Builder. The core product. A canvas where you compose real DAGs of nodes — LLM calls, conditional branches, HTTP requests, code nodes, tool nodes, knowledge-base lookups, variable assignments. Every non-technical PM we've watched touch it produces something running within an hour. Every engineer rolls their eyes at first and then admits, an hour later, that it saved them a day of scaffolding.
Native RAG (Knowledge Bases). Upload PDFs, docs, or scrape URLs. Dify handles chunking (multiple strategies including semantic), embedding (choose your model), vector storage, retrieval, and reranking. You can drop a knowledge-base node into any workflow and query it as a first-class step. This is the feature most teams underestimate before using Dify and then can't imagine building without.
Agent Mode. Beyond simple workflows, Dify supports true agentic loops — an LLM that plans, calls tools, observes results, and iterates until it reaches a goal. Configurable tool set, configurable model, configurable step limits. Not as sophisticated as CrewAI's multi-agent orchestration, but sufficient for most single-agent use cases.
Tool Ecosystem. 50+ prebuilt tools you can wire into workflows — Google Search, SerpAPI, WolframAlpha, Slack, DALL-E, Stable Diffusion, code interpreters, custom HTTP APIs, and native connectors to Zapier and Make. You can define your own tool via OpenAPI spec.
Model Router. Choose your LLM per node. OpenAI, Anthropic, Google, Azure, self-hosted Ollama, Groq, Fireworks — all first-class. You bring your own API keys, and Dify tracks token usage per model per workflow. Model switching is a config change, not a rewrite.
Prompt Playground and Versioning. Test prompts against multiple models side-by-side, save versions, roll back. Prompt engineering as a first-class product surface, not a hidden config file.
Chatflow and Workflow Modes. Two workflow flavors — Chatflow is optimized for conversational agents (chat history, sessions, streaming); Workflow is optimized for one-shot generation (form input, structured output). Same canvas, different runtime.
API and Embed. Every workflow you publish becomes a REST API endpoint automatically, complete with authentication, rate limiting, and monitoring. You can also embed workflows as chat widgets on any site — instant chatbot deployment.
Open-Source Self-Hosting. Dify's core is Apache-2.0 on GitHub. You can pull the Docker Compose file and run the whole platform on your own infra if compliance demands it or if you'd rather not pay the hosted fee. Real escape hatch.
Team Collaboration. Multi-user workspaces with roles, shared knowledge bases, and prompt libraries. Not as deep as an enterprise LLMOps platform, but sufficient for a team of 5-20.
Monitoring and Logs. Every workflow run is logged with inputs, outputs, model calls, latency, and cost. You can inspect, replay, and debug. This is table-stakes but many competitors still ship it poorly.
Dify offers a hosted plan structure alongside a free open-source self-host option.
Sandbox (Free) — Hosted, 200 messages/day. Full workflow builder, RAG, agents, all core features. Enough to prototype and validate.
Professional ($59/mo) — Hosted, higher message quota (~5,000/mo), 5 team seats, 200 knowledge documents, 3 apps.
Team ($159/mo) — Hosted, higher quota (~10,000/mo), 20 team seats, 500 knowledge documents, 50 apps, advanced logging retention.
Enterprise (custom) — SSO, SLAs, dedicated support, custom quotas.
Self-Hosted (Free / Apache-2.0) — Deploy the open-source Docker Compose stack yourself. No message limits, no team-size limits. Your only costs are infra and your own LLM API keys.
| Feature | Sandbox | Professional | Team | Self-Host |
|---|---|---|---|---|
| Messages/day | 200 | ~166 avg | ~333 avg | Unlimited |
| Team seats | 1 | 5 | 20 | Unlimited |
| Knowledge docs | Limited | 200 | 500 | Unlimited |
| Apps | 3 | 3 | 50 | Unlimited |
| Model providers | All | All | All | All |
| Advanced logs retention | 30 days | 30 days | 90 days | Your infra |
| SLA | No | No | No | No |
| Starting price | $0 | $59/mo | $159/mo | $0 + infra |
LLM inference costs are pass-through — you supply your own OpenAI / Anthropic / etc. API keys, and you pay those providers directly. Dify's fees are for the runtime, not the model tokens.
The self-hosted option is the escape hatch that makes the pricing sane: if the hosted quota doesn't fit your economics, you deploy the same code to your own AWS or GCP instance for the cost of a small server.
Pros.
Cons.
Product teams shipping customer-facing agents. The archetype. You're a startup building an AI-native product feature — support agent, research assistant, sales SDR bot. Dify gets you from spec to embedded chat widget on your site in a week.
Solo founders needing something in production by Friday. The reason we recommend Dify to solo builders in AIQORA's community. It's the fastest path from "I have an LLM idea" to "I have a deployed endpoint my app can call."
Non-technical PMs building internal tools. Dify's visual builder is legitimately usable by a product manager who's never touched Python. Internal Q&A bots, spec generators, meeting-note summarizers — all deployable without engineering time.
Teams evaluating RAG at scale. The native knowledge-base primitive plus multiple chunking strategies plus reranking makes Dify a genuinely good sandbox for RAG experimentation. Even if you eventually rebuild in LangChain for control, prototyping in Dify saves weeks.
Regulated or air-gapped environments. Self-hosted Dify with self-hosted Ollama endpoints is a real path for compliance-sensitive teams — healthcare, legal, government — that can't send prompts to OpenAI.
Bad fit. Teams whose competitive advantage is the runtime itself (build in LangGraph). Multi-agent orchestration with rich inter-agent communication (CrewAI). Pure prompt-tuning shops with no workflow complexity (any LLM playground plus a Python script).
LangChain / LangGraph. The code-first, deeply-flexible option. If you need custom retrieval logic, custom agent-state, or integrations Dify doesn't cover, LangChain gives you a Python SDK and expects you to build the runtime. Steeper learning curve, longer time-to-production, but no ceiling. Read the full comparison in Dify vs LangChain vs CrewAI: Which AI Agent Framework in 2026?.
CrewAI. The multi-agent orchestration specialist. If your problem is "researcher agent debates with editor agent who checks with fact-checker agent," CrewAI models that natively and Dify does not. Python-first, code-defined, opinionated about roles and goals. Full head-to-head in our agent framework showdown.
Flowise. The closest direct visual competitor — open-source, node-based, built on LangChain. Flowise is scrappier and less polished than Dify but 100% open-source with no hosted tier at all. If self-host-first is a hard requirement, Flowise is worth evaluating.
Honorable mentions: n8n (now with strong AI nodes, better if your core problem is automation with LLMs sprinkled in), Make (visual automation with AI integrations), and Voiceflow (specialized for conversational agents).
Getting a Dify workflow live takes about an hour for a first working agent. Sign up at https://affiliate.dify.ai/olk86803g6fe-nusw0l — Sandbox is free (200 messages/day) and gets you access to every core feature.
Once you're in: (1) Create a workspace, add your model provider keys — OpenAI, Anthropic, or self-hosted Ollama. (2) Create your first app: choose Chatflow (conversational) or Workflow (one-shot). (3) Build the simplest possible flow — user input → LLM node → response. Publish. Test the API endpoint. (4) Add a knowledge base — upload a few docs, wire in a knowledge-base node, watch the LLM answer questions with citations. (5) Add tools — HTTP request, SerpAPI, whatever your workflow needs. (6) Copy the API key and cURL the endpoint from your app.
If you're evaluating self-hosting, the docker-compose file on GitHub runs on any machine with Docker in under 5 minutes. Same platform, no message quota, your own infra costs.
Is Dify really open-source? Yes, Apache-2.0 licensed core on GitHub. You can self-host the full platform. There's a Dify Enterprise tier with proprietary add-ons (SSO, advanced compliance), but the core workflow engine is genuinely open.
How does Dify compare to LangChain? Dify is a platform; LangChain is a library. Dify is faster to production and has a lower ceiling. LangChain is slower to production and has no ceiling. For most product teams, start with Dify and graduate to LangChain only when the ceiling becomes a real problem. Full breakdown in our agent framework comparison.
Do I bring my own LLM keys? Yes. Dify's fees are for the runtime — the LLM tokens are pass-through to whichever provider you choose (OpenAI, Anthropic, Google, Groq, etc.). This means Dify doesn't mark up your model costs.
Can I use Dify for multi-agent workflows? Basic multi-agent works — you can call sub-workflows from a parent workflow. Rich inter-agent conversations (like CrewAI's role-based debate) do not. If that's your core need, use CrewAI.
How is RAG performance? Strong out of the box. Multiple chunking strategies (fixed, semantic, hierarchical), multiple embedding models, integrated reranking. Comparable to a well-configured hand-rolled LangChain RAG stack, at a fraction of the setup time.
What about production monitoring? Every workflow run is logged with inputs, outputs, model calls, latency, and cost. You can replay and debug from the logs. Not as deep as a dedicated LLMOps tool (Langfuse, Helicone), but sufficient for most teams.
How stable is the platform? In our production use since 2024, Dify has had a small number of brief outages but no data-loss events. Uptime is comparable to other mid-stage SaaS platforms in the LLMOps space.
Dify is the correct choice for product teams, solo founders, and non-technical PMs who need an AI workflow or agent in production this week. The visual builder plus native RAG plus 50+ tool integrations plus the API-endpoint-per-workflow pattern is the fastest path from "AI idea" to "endpoint the app can hit" in the current market. The escape hatch — Apache-2.0 self-host — protects you from vendor risk. That combination is unique in 2026.
If your competitive advantage is the runtime itself, use LangChain. If your problem is multi-agent orchestration, use CrewAI. For nearly every other case where you want a working agent shipped this week, use Dify.
Start on the Sandbox tier at https://affiliate.dify.ai/olk86803g6fe-nusw0l. 200 messages a day is enough to prove the workflow before paying. Upgrade to Professional at $59/mo when you have paying users; upgrade to Team at $159 when you have a team; self-host when hosted quotas stop making sense. That progression is genuinely as clean as it sounds.
AIQORA earns a commission if you sign up via our link — we test before we recommend.