How I Built My Own Automation Hub (And the Problems That Nearly Stopped Me)

How I Built My Own Automation Hub (And the Problems That Nearly Stopped Me)

I hit the wall with scripts on a Tuesday afternoon. One automation had become three, then six. Each one lived in its own file, had its own schedule, its own way of breaking. When something failed at 2am, I spent more time figuring out which script was the culprit than actually fixing it.

That was when I decided to try n8n. This is the real account of what I built, what went wrong on the way, and whether it was worth the afternoon I spent on it.

What I was trying to solve

I wanted a self-hosted hub where I could chain AI conversations to other triggers, pull data from websites automatically, and send results to my Telegram without managing a tangle of Python files. No per-task pricing, no vendor lock-in, everything running on my own server.

Prerequisites are minimal: a Linux server with Docker already installed, a DeepSeek API key, and about two hours of focused time.

Part 1: The installation that almost did not happen

I had Docker on the machine already, but if you are starting fresh, the first step is:

yum install -y docker-ce docker-ce-cli containerd.io
systemctl start docker
systemctl enable docker

Then I ran into the first real problem. The Docker image pull for n8n kept timing out. I was on a Chinese cloud provider and Docker Hub pulls were painfully slow. The fix was configuring registry mirrors before trying to pull anything. I edited /etc/docker/daemon.json with mirror sources, reloaded Docker with systemctl restart docker, and the pull went through without a hitch. This step is easy to skip if you do not know it is coming.

Once the image was pulled, I set up the data directory:

mkdir -p ~/.n8n
chmod -R 777 ~/.n8n

Then the container:

docker run -d --name n8n -p 5678:5678 -v ~/.n8n:/home/node/.n8n -e N8N_BASIC_AUTH_ACTIVE=true -e N8N_BASIC_AUTH_USER=admin -e N8N_BASIC_AUTH_PASSWORD=YourStrongPass123! -e TZ=Asia/Shanghai -e N8N_COOKIE_SECURE=false -e N8N_SECURE_COOKIE=false --restart always n8nio/n8n:latest

I pointed my browser at http://server-ip:5678, logged in with the credentials, and the dashboard was there. First time setup took about ten minutes after I cleared the mirror issue.

Part 2: Connecting DeepSeek was the easy part

Once inside, I created a new workflow and added an AI Agent node. Selected DeepSeek from the model dropdown, pasted in my API key, and connected it to a manual trigger. DeepSeek has a free tier that is enough to test with, so I did not need to spend anything to get started.

Sent a test message. Got a coherent response back. The AI-powered workflow was running.

What surprised me was how straightforward the whole chain was. Trigger -> AI Agent -> output. The n8n canvas shows you the flow visually, which is much easier to read than six lines of Python doing the same thing.

Part 3: Four problems I actually hit

Problem 1: The image pull timeout

I already mentioned this. If you are in a region with slow Docker Hub access, configure mirrors in /etc/docker/daemon.json before you run docker pull. Not after. It is the first thing to check if your pull fails.

Problem 2: Authentication broke after first login

After I created the account, n8n kept asking me to log in again. My credentials were correct, but something was off with the session. It turned out I had missed setting N8N_COOKIE_SECURE=false in the docker run command. Without it, n8n was rejecting its own cookies. Adding those two environment variables fixed it immediately.

Problem 3: The workflow ran but returned nothing

I built a second workflow for web scraping. It appeared to execute without errors, but produced no output. The canvas had no red nodes, which initially made me think nothing was wrong. Then I noticed the AI Agent node had no output either. I clicked into it and found it had silently failed because my prompt was too vague. Tightened the prompt, the workflow ran correctly. The lesson: in n8n, no red node does not always mean no problem. Inspect each node’s output individually.

Problem 4: I could not access it from outside my network

The n8n dashboard only worked from within the server’s network. I wanted to trigger workflows from my phone and receive webhook callbacks from third-party services. The fix I went with was cpolar, which creates an HTTPS tunnel to your local port with one command and gives you a public URL. I am aware this opens port 5678 to the internet, so I will address the security question in the suggestions section below.

Part 4: What I would do differently

Suggestion 1: Build one workflow and make it solid before adding more

I wanted to build five workflows on day one. I forced myself to stop after two. The debugging tools, the retry mechanics, the data transformation nodes . you encounter them properly only when you are maintaining a workflow over time, not when you are rushing to build everything at once.

Suggestion 2: Set up error notifications before you need them

By default, n8n does not tell you when a workflow fails. I added an error trigger to each workflow that sends a message to my Telegram channel the moment something breaks. Took five minutes. Now I find out about failures before anyone else does.

Suggestion 3: Build the web scraping workflow before you actually need it

I built the scraper after I already had a urgent need for it at 11pm. The workflow itself is simple . HTTP Request node to fetch the page, HTML node to extract the content, action node to send it somewhere. Having it ready in advance means you just hit run instead of scrambling to build it under pressure.

Suggestion 4: cpolar is fine for personal use, not for enterprise

cpolar solved my access problem cleanly. But I am clear-eyed about what it is doing: it exposes port 5678 to the public internet. For a personal homelab, that is an acceptable trade-off. For anything beyond that, I would put n8n behind a reverse proxy with proper TLS, authentication, and IP allowlisting. The cpolar approach works until it does not.

Suggestion 5: Use the canvas as your documentation

Every time I finish a workflow, I save it with a descriptive name and add a note to the canvas. Three months from now, opening the workflow tells me exactly what it does and where it might be failing. This sounds obvious but it is the single biggest quality-of-life improvement over script files, where you have to re-read the code to remember what you were thinking.

The honest assessment after a week

Two workflows running daily. The AI chat workflow logs responses to a Notion database on a schedule. The web scraping workflow sends results to Telegram on demand. Both have been more reliable than the scripts they replaced.

The setup took about two hours, most of which was reading documentation. The actual implementation was faster than I expected. The biggest surprise was how much more I automate now that the infrastructure is already in place. The cost of building a new workflow dropped from writing a script to drawing a flowchart. That changes what you are willing to automate.

If you are managing automation with scripts and bash chains, n8n is worth a Saturday afternoon. The learning curve is shallow and the capabilities are broad. You do not need to commit to it permanently to get real value from it. Even one workflow that saves you twenty minutes a day pays back the setup time in two weeks.