What do 7,698 real-world automation workflows have in common?
This project extracts the universal patterns — and turns them into explorable blueprints.
This project parses 7,698 n8n workflow JSONs, maps 80+ node types to 9 abstract categories, and extracts the recurring patterns behind everyday automation. The result: an interactive explorer with blueprints for 12 use cases, and an MCP server that lets Claude Code query the pattern database directly.
People build automation workflows every day — in n8n, Make, Zapier, or with code. The tools differ, but the patterns are universal:
trigger -> ai -> deliver (AI writes content, sends it somewhere)
trigger -> transform -> data (data arrives, gets cleaned, gets stored)
trigger -> api -> logic -> api (two systems talk to each other)
These patterns repeat across thousands of workflows. If you know the pattern, you know the architecture — regardless of which tools you use.
This project extracts those patterns from real data and makes them actionable.
git clone https://github.com/janrummel/workflow-patterns
cd workflow-patterns
uv sync
# Generate the interactive site
uv run python scripts/generate_site.py
open docs/index.html
# Or use as MCP server with Claude Code
claude --mcp-config mcp.json| Insight | Detail |
|---|---|
| 7,698 workflows analyzed | From the n8nworkflows.xyz public archive |
| 2,592 unique patterns | Distinct category sequences found across all workflows |
| 12 use case clusters | AI Chatbot, Content Creation, Data Pipeline, Email Automation, and more |
| Most common pattern | trigger -> transform -> api -> logic (187 workflows) |
| AI is mainstream | 45%+ of workflows include at least one AI node |
| Most workflows are small | 4-6 nodes is the sweet spot — you don't need 20 steps |
The site lets you pick a use case and instantly get a complete blueprint:
- The pattern as an abstract step sequence
- Building blocks needed to implement it
- Tools & services grouped by category — showing alternatives, not just the most popular
- Starter code in Python with syntax highlighting
- Claude Code architecture recommendation
Generate the interactive site and open it in your browser:
uv run python scripts/generate_site.py
open docs/index.htmln8n JSON files (7,698 workflows)
|
v parse
Workflow objects (nodes + edges)
| handles inconsistent JSON formats: null connections,
| string-wrapped lists, missing fields
v categorize
Abstract graphs (80+ node types -> 9 categories)
| "lmChatOpenAi" -> ai
| "googleSheets" -> data
| "gmail" -> deliver
v analyze
Patterns, statistics, use case clusters
| BFS graph traversal for pattern extraction
| Jaccard similarity for workflow matching
v translate
Claude Code architectures
| pattern -> MCP servers + agents + scripts
v serve
Interactive site + MCP server for Claude Code
Every node type maps to one abstract category. This is what makes patterns universal — the specific tool doesn't matter, only its role in the workflow.
| Category | What it represents | Example nodes | Claude Code equivalent |
|---|---|---|---|
| trigger | Starts the workflow | scheduleTrigger, webhook, formTrigger | Cron job, file watcher |
| ai | LLM processing | openAi, lmChatOpenAi, agent | Claude agent |
| transform | Reshape data | set, code, filter, aggregate | Python script |
| deliver | Send output | gmail, slack, telegram | MCP server |
| data | Read/write storage | googleSheets, postgres, airtable | MCP server |
| api | Call external APIs | httpRequest | MCP server |
| logic | Branch, route, merge | if, switch, merge | Claude skill |
| storage | File operations | googleDrive, s3 | MCP server |
The project includes an MCP server that exposes pattern analysis to Claude Code. Connect it and ask natural language questions about automation patterns.
claude --mcp-config mcp.jsonThen ask Claude:
"I want to build a workflow that fetches news articles weekly, summarizes them with AI, and sends a digest by email."
Claude will search the pattern database, find matching workflows, and suggest an architecture:
**Query:** Send a weekly AI summary of news articles by email
**Detected categories:** trigger, ai, deliver
**Matching workflows:**
- [80% match] Weekly AI Newsletter
Pattern: trigger -> ai -> transform -> deliver
- [60% match] RSS Digest Bot
Pattern: trigger -> api -> ai -> deliver
**Suggested pattern:** trigger -> ai -> deliver
| Tool | What it does |
|---|---|
search_patterns(query) |
Find patterns matching a natural language description |
suggest_implementation(pattern) |
Get Claude Code architecture for a pattern |
list_categories() |
Show category frequency and common connections |
show_all_patterns() |
List all unique patterns ranked by frequency |
Each example implements a real workflow pattern with current tools — not just a blueprint, but working code you can run.
| Example | Pattern | What it does |
|---|---|---|
| AI Content Creation | api → ai → transform → deliver |
Fetches RSS feeds, summarizes with Claude, saves a markdown digest |
| AI Chatbot | trigger → ai → data → deliver |
Streaming terminal chat with 5 persona presets and conversation persistence |
| Email Automation | trigger → transform → deliver |
Template-based emails with 5 presets, HTML rendering, and optional SMTP delivery |
| Scheduled Notifications | trigger → api → logic → deliver |
Rule-based notification system with urgency levels and digest output |
| Data Pipeline | trigger → data → transform → data |
Data processing with filter, aggregate, sort, rename, deduplicate transforms |
| API Integration | trigger → api → transform → data |
REST API data collector with JSON path extraction and sample fallback |
cd examples/ai-content-creation
uv sync
cp .env.example .env # add your ANTHROPIC_API_KEY
uv run python run.py # interactive feed selection
uv run python run.py --feeds https://hnrss.org/newest?points=100 # or customMore examples coming: Form/Webhook Handler, Social Media Automation, and more.
uv run pytest -v # 97 tests across 20 modules
uv run ruff check . # Lintworkflow-patterns/
├── src/workflow_patterns/
│ ├── models.py # Node, Edge, Workflow, Pattern + 80+ node type mappings
│ ├── parser/parse.py # Robust n8n JSON parser (handles format inconsistencies)
│ ├── patterns/analyzer.py # Pattern extraction, node stats, Jaccard similarity
│ ├── translator/claude_code.py # Pattern -> Claude Code architecture mapping
│ └── mcp_server/server.py # 4 MCP tools for Claude Code integration
├── examples/
│ ├── ai-chatbot/ # 22 tests — streaming API, 5 personas
│ ├── ai-content-creation/ # 30 tests — Claude API, 60 curated feeds
│ ├── email-automation/ # 25 tests — 5 templates, HTML + SMTP
│ ├── scheduled-notifications/ # 60 tests — rule engine, urgency levels
│ ├── data-pipeline/ # 53 tests — filter, aggregate, sort
│ └── api-integration/ # 36 tests — REST client, sample fallback
├── scripts/
│ └── generate_site.py # Static site generator with interactive wizard
├── docs/
│ └── index.html # Generated GitHub Pages site
├── tests/ # 20 unit tests across 4 modules
├── evals/ # 10 evaluation test cases
├── data/sample_workflows/ # 15 sample n8n workflows for development
├── .github/workflows/ci.yml # CI: lint + test on push/PR
└── mcp.json # MCP server config for Claude Code
Workflow data from n8nworkflows.xyz, an independent archive of 7,000+ public n8n workflow templates. The sample dataset includes 15 workflows for development; the full dataset (7,698 workflows) can be placed in data/all_workflows/.