Self-Host n8n with Local Ollama: Zero-Cost AI Workflows (2026)

Self-Host n8n with Local Ollama: Zero-Cost AI Workflows (2026)

I've been running n8n + Ollama locally for a few months. No OpenAI bill. No data leaving my network. The setup takes under 20 minutes — except for the Docker networking part, which cost me an hour the first time. This post skips the fluff.

Table of Contents

• What n8n and Ollama Do
• Hardware Checklist
• Install Ollama and Pull a Model
• Run n8n with Docker Compose
• Connect n8n to Ollama
• Fix Docker Networking
• 5 Workflows Worth Building
• Troubleshooting Table
• Local vs Cloud
• FAQ

What n8n and Ollama Do

n8n is a self-hostable workflow automation tool — Zapier, but you own the server. It connects triggers, logic, APIs, databases, and AI models via a visual node editor. Refer to the n8n Docs for official workflow node definitions.

Ollama is a local LLM runtime. It downloads model weights, loads them into memory, and exposes a REST API on port 11434. Any app that makes HTTP requests can talk to it. Grab the installer from the Ollama Download page.

Trigger → n8n → Ollama API (:11434) → Local LLM → Output

Real reasons to use this stack: no per-token cost, data stays local, works offline. The honest tradeoff: a 7B model on an M2 MacBook is slower and weaker than GPT-4o. Use local where privacy or cost matters; use cloud where quality ships to users.

Related Reading: When to Use Cloud LLMs Instead (Claude Code CLI vs Cursor)

Hardware Checklist

RAM	What Runs
8 GB	1B–3B models (llama3.2:3b, qwen2.5:3b)
16 GB	7B models comfortably (llama3.1:8b, qwen2.5:7b)
32 GB+	13B–30B; 70B+ needs a GPU on top

SSD is non-negotiable — models are 2–40 GB. Apple Silicon is great here: GPU and CPU share the same memory pool, so a 16 GB M2 runs 7B models faster than most 16 GB Intel machines.

Install Ollama and Pull a Model

macOS / Linux:

curl -fsSL https://ollama.com/install.sh | sh

Windows: grab the installer at ollama.com/download.

Confirm it's alive:

curl http://localhost:11434
# Returns: "Ollama is running"

Pull a model:

ollama pull llama3.2       # 3B — good for 8 GB
ollama pull qwen2.5:7b     # 7B — better quality, needs 16 GB
ollama pull codellama:7b   # Code-specific tasks

Quick test:

ollama run llama3.2   # Type a prompt. /bye to exit.
curl http://localhost:11434/api/tags   # Lists your pulled models

✅ Ollama running | ✅ Model pulled | ✅ API accessible

Read: Claude code CLI setup - Terminal AI Agent vs Cursor IDE

Run n8n with Docker Compose

I run Ollama directly on the host for GPU access. n8n lives in Docker. Here's my working docker-compose.yml:

version: '3.8'
services:
  n8n:
    image: docker.n8n.io/n8nio/n8n
    ports:
      - "5678:5678"
    volumes:
      - n8n_data:/home/node/.n8n
    extra_hosts:
      - "host.docker.internal:host-gateway"  # Linux only

volumes:
  n8n_data:

docker compose up -d

Open http://localhost:5678, create your account (local, no cloud), and you're in.

✅ n8n running

Read: Open-Weight Reasoning Models vs. Proprietary APIs: Cost, Latency & Privacy Benchmark

Connect n8n to Ollama

Go to Settings → Credentials → Add Credential → Ollama.

The Base URL field is where people get stuck. Use this lookup matrix:

n8n Location	Ollama Location	Base URL
Host machine	Host machine	http://localhost:11434
Docker	Host machine	http://host.docker.internal:11434
Docker	Docker (same network)	http://ollama:11434

Click Test. Green = connected. ✅

Fix Docker Networking

This is where Docker breaks most setups. Here's why: inside a container, localhost refers to the container itself — not your machine. So localhost:11434 looks for Ollama inside the container, where nothing is listening.

• macOS / Windows (Docker Desktop): host.docker.internal resolves automatically. Just use it.
• Linux: It doesn't exist by default. The extra_hosts line in the compose file above handles it. Check the Docker extra_hosts documentation for configuration syntax.

--add-host=host.docker.internal:host-gateway

If running both containers in Docker, put them on the same network and reference Ollama by service name (http://ollama:11434). Note that GPU passthrough in this scenario requires the NVIDIA Container Toolkit on Linux.

5 Workflows Worth Building

1. Email summarizer: Gmail trigger → Ollama → log summary to Google Sheets
2. Document summarizer: Drive PDF → extract text → Ollama → save back to Drive
3. Content draft generator: Webhook (topic input) → Ollama → Notion
4. RAG / knowledge base: Add Qdrant. Explore the pre-built n8n Self-Hosted AI Starter Kit to launch n8n + Ollama + Qdrant + PostgreSQL automatically.
5. Private chat assistant: Telegram/Webhook → Ollama with custom system prompt → reply. Nothing touches a cloud API.

Troubleshooting Table

Error	Cause	Fix
Connection refused	Ollama not running	ollama serve
localhost fails	n8n is in Docker	Use host.docker.internal:11434
host.docker.internal fails on Linux	Missing mapping	Add extra_hosts: host-gateway
Model not found	Not pulled	ollama pull <model>
Credential passes, workflow fails	Name mismatch	Run ollama list — names must match exactly
Slow responses	Model too large	Switch to Q4-quantized variant
OOM errors	Not enough RAM	Drop to 3B model or add swap
Workflows lost on restart	No volume	Mount n8n_data (in compose above)

Local vs Cloud

Factor	Ollama + n8n	Cloud API
Cost per call	Zero	Pay per token
Data privacy	Your machine	Provider's servers
Offline use	Yes (after download)	No
Setup effort	Higher	Minimal
Model quality	7B–70B local	Frontier models available
Scaling	Your hardware ceiling	Effectively unlimited

FAQ

Can I run n8n and Ollama for free?
Yes. n8n self-hosted is free under its fair-code license. Ollama is open source. Real costs are electricity and your hardware.

Open Source Cursor Alt for CLI

What URL does n8n use to reach Ollama?
Host-to-host: http://localhost:11434. Docker-to-host: http://host.docker.internal:11434.

Why does localhost fail inside Docker?
localhost inside a container is the container itself, not your host machine. Ollama runs on the host, creating a network mismatch. Use host.docker.internal.

Which model should I start with?
8 GB RAM → llama3.2:3b. 16 GB → qwen2.5:7b. Pick what fits in your free RAM after your OS takes its share.

Does n8n support Ollama natively?
Yes. Built-in credentials and an Ollama Chat Model node are available out of the box.