Dify: How to Build Your First AI Workflow Without Writing Infrastructure
If you ever wanted to connect AI to a real flow — pull data from an API, pass it through a prompt, route the result to another system — you probably hit the same wall: lots of boilerplate before anything useful happens.
Dify is the answer to that wall. It’s an open-source platform for building production-grade AI workflows visually, without rebuilding the infrastructure stack every time. 130,000 stars on GitHub and a $30M funding round in March 2026 confirm that the dev community found it genuinely useful.
Let’s build something real with it.
What exactly is Dify?
Think of it as a visual IDE for AI applications. Instead of writing code to handle prompt templates, API calls, context management, RAG pipelines, and observability — you drag and connect nodes on a canvas.
It occupies a specific spot in the ecosystem: it’s not a code agent (that’s Claude Code or Cursor territory), and it’s not an enterprise platform that takes months to set up. It’s the middle layer — where teams turn their domain knowledge into AI workflows they can maintain in production.
Core capabilities:
- Visual Workflow Builder — drag-and-drop canvas for multi-step AI pipelines
- Model-agnostic — supports Claude, GPT, Gemini, Mistral, Llama and any OpenAI-compatible API
- RAG Pipeline — document ingestion, chunking and retrieval, all included
- Agent Framework — 50+ built-in tools: web search, code execution, DALL·E and more
- MCP Support — connect external APIs via MCP, or publish your Dify workflow as an MCP server
- LLMOps — logs, performance monitoring and prompt iteration in production
Runs in the cloud (cloud.dify.ai) or self-hosted with Docker.
Setup in 3 minutes
# Option 1: Cloud (no setup)
# Go to cloud.dify.ai → create free account → start building
# Option 2: Self-hosted with Docker
git clone https://github.com/langgenius/dify.git
cd dify/docker
cp .env.example .env
docker compose up -d
# Access at http://localhost/
For your first workflow, the cloud is perfect. Self-hosting makes sense when you need data privacy or want to run local models.
Once inside, go to Studio → Create App → Workflow.
Building your first workflow: a practical example
Let’s build something concrete: a workflow that takes a GitHub repository URL, fetches the README and generates a structured summary (what it does, who it’s for, how to get started) — ready to post in a Slack message.
Step 1: Start Node
Every workflow starts with a Start node. Here you define your input variables:
Input variable: repo_url (string)
Example: https://github.com/langgenius/dify
Step 2: HTTP Request Node
Connect an HTTP Request node to fetch the README:
Method: GET
URL: https://api.github.com/repos/{{repo_owner}}/{{repo_name}}/readme
Headers:
Accept: application/vnd.github.raw
You’ll need a small Code node before this to parse the owner and repo name from the URL — about 3 lines of Python.
Step 3: LLM Node
This is where the AI work happens. Connect an LLM node:
Model: claude-sonnet-4-20250514 (or whatever model you've configured)
System prompt:
You are a technical writer. Given a repository README, extract:
1. What the project does (1 sentence)
2. Who it's for (1 sentence)
3. How to get started (max 2-3 steps)
Respond as structured text, without markdown headers.
User: {{http_response.body}}
Step 4: End Node
Connect an End node with the LLM output as the response. Now you have a functional API endpoint you can call from anywhere.
# Test it from Dify's integrated runner, or call it as an API:
curl -X POST https://api.dify.ai/v1/workflows/run \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"inputs": {"repo_url": "https://github.com/langgenius/dify"}}'
With these four nodes you already have a functional workflow connected to a real endpoint. From here, you can add branching logic, additional LLM nodes, or integrations with other systems.
The MCP integration: publish your workflow as a tool
This is where Dify gets interesting for those already in the Claude Code / MCP ecosystem. You can take any workflow you build and expose it as a standard MCP server — invocable from Claude Code, Cursor, or any MCP client.
In Dify: Settings → MCP Server → Enable. You get a server URL you can put directly in your claude_desktop_config.json or Claude Code’s MCP config.
What this means: you build a complex workflow once in Dify’s visual editor (with RAG, multi-step logic, observability), and your code agent can invoke it as a tool. No glue code.
The Human Input node (v1.13.0)
One of Dify’s most useful recent additions: a Human Input node that pauses a workflow and waits for a human decision before continuing.
This matters for any agentic workflow where you want guardrails — approval flows, content review, exception handling. The workflow runs autonomously until it hits something that needs a human decision, then resumes after input.
Workflow: ingest support ticket →
LLM classifies severity →
If severity = critical → Human Input node (approve escalation?) →
If approved → trigger PagerDuty alert
Pricing
| Plan | Price | For whom |
|---|---|---|
| Sandbox (Free) | $0 | Testing, up to 200 message credits |
| Professional | $59/month | Small teams, production workloads |
| Team | $159/month | Up to 25 workspace members |
| Enterprise | Contact sales | Compliance, self-hosted, SLA |
The self-hosted Community Edition is completely free with no limits beyond your own infrastructure.
When does it make sense to use Dify?
Good choice if:
- You need to connect AI to existing business systems (databases, APIs, internal tools) without building a custom backend
- You’re building RAG applications and want the pipeline managed
- Your team includes non-technical people who need to iterate on prompts and workflows
- You want observability over AI calls in production
Not the right tool if:
- Tasks are purely code — that’s Claude Code territory
- It’s a single-prompt use case — direct API calls are simpler
- You need very low latency and maximum control — raw API access wins
Conclusion
Dify fills a real gap: it’s where AI connects to your existing systems, not just your terminal. The visual workflow builder lowers the barrier for teams that need to launch AI automation without a dedicated ML engineer. MCP support means it works well with the tools the dev community already uses.
130K stars don’t happen by accident.
Are you already using any AI workflow orchestration tools? Did you build it yourself or are you using something like Dify or n8n? Let us know in the comments 