This guide explains a self-hosted n8n content pipeline with separate research, drafting, validation and human review stages. Self-hosting describes the orchestrator; connected model and search services may still be remote.
Visual guide: Build a publishing pipeline around a review gate
How to use this map
- Use a stable task identifier and preserve source URLs with research notes. A source count alone does not establish that a claim is supported.
- Make the writer produce fields your CMS expects. Separate title, body, description, labels and image information rather than passing an opaque text blob.
- Validate types and nonempty fields, then review factual claims and image rights. A parser can accept structurally valid content that is still unusable or unsupported.
- Make the first CMS operation a draft and keep its returned identifier. Before retrying, check whether that draft already exists so a network interruption does not create duplicates.
Stop and check: Self-hosting n8n does not remove provider quotas or make automatic publication trustworthy. Public release remains a separate editorial decision.
Next practical guides: n8n AI Agent with Gemini and Hugging Face MCP: A Real Test · Build a Local AI Workflow with Ollama, ComfyUI and n8n
Why n8n Won the Self-Hosted Automation Crowd
You own it. Self-host with Docker; your API keys, prompts, and data stay on your infrastructure. For anyone handling client or proprietary content, that alone is decisive.
It speaks HTTP fluently. Beyond hundreds of prebuilt integrations, the generic HTTP Request node means any REST API — including your local ComfyUI or Ollama server — is reachable.
Real logic, not just linear steps. Branching (IF/Switch), loops, merges, error handling, and scheduled or webhook triggers make it a genuine orchestration layer, not a toy.
AI-native nodes. Built-in nodes for LLM chat, agents, and vector stores mean you are not hand-rolling every model call.
Step 1: Self-Host with Docker
The fastest path is Docker. A minimal setup:
docker volume create n8n_data
docker run -it --rm --name n8n -p 5678:5678 -v n8n_data:/home/node/.n8n docker.n8n.io/n8nio/n8n
Open http://localhost:5678, create your owner account, and you have a working instance. For anything persistent, move to a docker-compose file with a database and set the environment variables for your domain, timezone, and encryption key. Put it behind a reverse proxy (Caddy or Nginx) with HTTPS before exposing it.
Step 2: Design the Pipeline
Let's build a concrete, useful flow: on a schedule, draft a blog section with an LLM, generate a matching image locally, and post a draft.
Schedule Trigger — fire daily, or use a Webhook node so an external event kicks it off.
LLM node (or HTTP Request to your model) — send a prompt like "Write a 150-word intro on today's topic: {{topic}}." Map the output to a variable.
HTTP Request to ComfyUI — POST a workflow to your local ComfyUI's /prompt endpoint to generate a featured image (see our ComfyUI API guide for the payload shape). Poll for completion, then fetch the image.
IF node — branch on quality or length checks; route failures to a notification instead of publishing.
Publish node — push to your CMS, a Google Doc, Notion, or send to a review channel. For a blog, a draft-status post is the safe default.
Passing Data Between Nodes
n8n moves data as JSON items between nodes. You reference upstream output with expressions like {{ $json.text }} or {{ $node["LLM"].json.output }}. Spend ten minutes understanding the item model — most beginner friction is here. The Edit Fields (Set) node is your friend for reshaping data between steps.
Step 3: Add Robustness
A demo that works once is not a pipeline. Add:
Error Trigger workflow — a separate flow that runs whenever any workflow errors, so failures reach you instead of failing silently.
Retry settings on flaky HTTP nodes (image generation and remote APIs especially).
A human-in-the-loop gate — post to a review channel and require approval before anything publishes publicly. For AI-generated content this is not optional.
Rate/cost guards — cap how many items a run processes so a bad trigger cannot fire 500 model calls.
Optimization and Operational Tips
Keep model calls local where you can. Pointing the HTTP node at a local Ollama or ComfyUI server removes per-call cost and keeps data in-house.
Modularize. Use Execute Workflow to call sub-workflows (e.g., a reusable "generate image" flow) instead of copy-pasting node clusters.
Version your workflows. Export the workflow JSON and commit it to git; n8n workflows are just JSON, which makes them reviewable and revertible.
Watch execution logs. The execution list shows exactly what data each node received — it is the single best debugging tool.
Mind the license. n8n is source-available under a fair-code (Sustainable Use) license, not classic open source — fine for internal automation and most business use, but read the terms before building a commercial hosted product on top of it.
Conclusion
n8n turns a pile of API calls into an owned, observable, and robust pipeline. Start with Docker, build one honest end-to-end flow (trigger to reviewed output), keep model inference local when you can, and add error handling and a human gate before you let anything publish. It is the connective tissue that makes the models in the rest of this series into an actual product workflow rather than a collection of one-off scripts.
Downloadable Workflow & References
|
