Self-Host n8n and Ship a Production AI Ops Agent

By the end of this playbook you will have:

• A self-hosted n8n instance on a VPS, reachable only over HTTPS on your own subdomain
• Credentials stored in n8n (not in chat logs), with a durable encryption key backed up offline
• One working AI Ops Agent workflow: Slack (or Telegram) question → agent uses tools → posts a useful ops result (CRM lookup or Slack summary)
• A production harden pass: secrets, HTTPS, backups, permissions, and a human approval gate before any CRM write

Gather these before you start:

1. VPS — Ubuntu 22.04 or 24.04, 2 GB RAM minimum (4 GB comfortable), 20+ GB disk, public IPv4. Hetzner, DigitalOcean, Contabo, Linode all work.
2. Domain you control — you will create n8n.yourdomain.com (A record → VPS IP).
3. SSH access as a sudo user.
4. Docker Engine + Docker Compose plugin on the VPS (install steps below if missing).
5. Accounts for the first business outcome: Slack workspace (or Telegram bot) where ops lives; CRM you already use (HubSpot, Pipedrive, or Salesforce) or a Google Sheet stand-in for week one; an LLM API key (OpenAI, Anthropic, or another provider n8n’s LLM nodes support).
6. A password manager — you will store: VPS root/user password, n8n owner password, N8N_ENCRYPTION_KEY, DB password (if you add Postgres later), API keys.

1. In your DNS provider, create an A record:

• Name: n8n (or the subdomain you prefer)
• Value: your VPS public IPv4
• TTL: 300 (or Auto)

1. Wait until dig +short n8n.yourdomain.com returns the VPS IP from your laptop.
2. Open ports 80 and 443 on the VPS firewall (uFW / cloud security group). Do not expose 5678 to the world — n8n stays on localhost; Traefik publishes HTTPS.

SSH in, then:

sudo apt update && sudo apt install -y ca-certificates curl
# Follow Docker’s current Engine install for Ubuntu, then:
sudo apt install -y docker-ce docker-ce-cli containerd.io docker-compose-plugin
docker --version
docker compose version

Optional: add your user to the docker group so you can run compose without sudo every time (usermod -aG docker $USER, then re-login).

n8n’s current production-shaped Docker Compose path uses Traefik for TLS and binds n8n to 127.0.0.1:5678. As of late September 2026, docs list stable ~2.40.x — pin or use the documented n8nio/n8n image tag your ops standard allows; prefer stable over beta.

mkdir -p ~/n8n-compose/local-files
cd ~/n8n-compose

Create .env (edit values):

DOMAIN_NAME=yourdomain.com
SUBDOMAIN=n8n
GENERIC_TIMEZONE=Asia/Karachi
SSL_EMAIL=you@yourdomain.com

Create compose.yaml from n8n’s Docker Compose guide — Traefik + n8n services, volumes n8n_data and traefik_data, WEBHOOK_URL=https://${SUBDOMAIN}.${DOMAIN_NAME}/, N8N_PROTOCOL=https, port bind 127.0.0.1:5678:5678.

Add an encryption key before first start (critical):

# Generate once; store in password manager AND in .env
openssl rand -hex 32

Add to .env:

N8N_ENCRYPTION_KEY=paste_the_hex_here

And under the n8n service environment: in compose:

- N8N_ENCRYPTION_KEY=${N8N_ENCRYPTION_KEY}

Start:

docker compose up -d
docker compose ps

Open https://n8n.yourdomain.com. Create the owner account (first user = admin). Save that password in your vault.

In n8n:

1. Settings → Users — invite only people who should edit workflows. Prefer SSO later; for day one, strong unique passwords.
2. Confirm the instance URL and webhook URL match HTTPS (Settings / environment already set WEBHOOK_URL).
3. Enable execution pruning (Settings → Executions) so disk does not fill with old runs — e.g. keep successful executions 7–14 days, errors longer if you triage.
4. Do not paste API keys into sticky notes or Slack. Use Credentials objects only.

LLM credential

1. Credentials → add your provider (OpenAI / Anthropic / etc.).
2. Name it clearly: prod-openai-ops — never test.
3. Restrict who can use the credential if your plan/RBAC allows.

Slack

1. Create a Slack app (or use n8n’s Slack credential OAuth flow).
2. Scopes typically needed for ops bots: post messages, read channels you choose, and (if using slash commands/events) event subscriptions pointing at your HTTPS webhook URL.
3. Install to the workspace; invite the bot to #ops or #revops.
4. Save the Slack credential in n8n.

Telegram alternative: create a bot via BotFather, paste the token into n8n’s Telegram credential, and use a Telegram Trigger.

Goal: an ops person asks in Slack: “What’s the stage and owner for deal Acme?” The agent looks it up and replies in-thread. No writes yet.

Workflow shape

1. Trigger — Slack Trigger (app mention or slash command) or Chat Trigger for a quick smoke test.
2. AI Agent node — system message that states:

• You are an internal ops assistant for COMPANY.
• You may read CRM / sheets tools.
• You must not invent deal stages; if lookup fails, say so.
• Keep answers short; cite the record id.

1. Chat Model — attach your LLM credential.
2. Memory (optional) — Window Buffer Memory for short threads.
3. Tools (read-only first):

• HubSpot / Pipedrive / Salesforce Get or Search nodes as Agent tools, or
• Google Sheets tool limited to a “Deals mirror” tab if CRM OAuth is blocked this week.

1. Reply — Slack node posts the agent’s final text back to the thread.

Smoke test

1. Activate the workflow.
2. From Slack, mention the bot with a real deal name that exists.
3. Confirm: correct stage/owner, no hallucinated fields.
4. In n8n Executions, open the run — every tool call should show real parameters.

When you are ready for writes (e.g. “Log that Acme asked for a quote”):

1. Add a CRM Update / Create note tool.
2. In the Agent’s tool connector, add a Human review (HITL) step for that write tool (Slack / Gmail / Teams / n8n Chat — see n8n’s Human-in-the-loop for tools docs).
3. Wire: Agent → HITL review → Write tool on the ai_tool path. The agent still “sees” the write tool’s name/schema; the review node intercepts execution.
4. Approval message must show actual parameters (deal id, fields to change) — not a model paraphrase.
5. Set limitWaitTime explicitly (e.g. 1–4 hours in seconds). Do not leave unbounded waits in production queues.
6. Deny path: agent should acknowledge “update not applied” and suggest a human follow-up.

Run this checklist before you call the agent “production”:

• [ ] https://n8n.yourdomain.com loads with valid cert
• [ ] Owner MFA/strong password stored in vault
• [ ] N8N_ENCRYPTION_KEY backed up offline
• [ ] Slack/Telegram bot responds in the intended channel only
• [ ] Read-only CRM (or Sheet) lookup succeeds on 3 real records
• [ ] Write tool blocked until HITL approve; deny leaves CRM unchanged
• [ ] Backup job ran once and a file exists off-box
• [ ] At least one other human can open Executions and revoke credentials
• [ ] Runbook link pasted in #ops (this playbook URL + who owns the VPS)

• Exposing port 5678 — bots and scanners will find it. Always reverse-proxy; bind localhost only.
• Losing the encryption key — credentials become permanent garbage. Backup the key the same day you create it.
• SQLite on a tiny disk with no pruning — disk full → silent outage. Prune executions and monitor disk.
• Webhook URL still http:// — Slack events fail randomly. Force HTTPS + matching WEBHOOK_URL.
• Agent writes without HITL — one bad tool call updates the wrong deal. Gate writes from day one.
• Shared “god mode” CRM token — use a dedicated integration user with minimum scopes.
• Beta image in production — prefer documented stable; schedule upgrades.

When this agent is live and gated, deepen approvals with Production Approvals: n8n Human-in-the-Loop Tool Gates, or connect a chat brain on top via OpenClaw + n8n over MCP.