n8n + Ollama: automate with 100% AI locale

#Why pair n8n with Ollama?

n8n is a visual orchestrator: you connect nodes (trigger, HTTP, database, LLM…) without writing code, and the workflow runs in the background on a cron job, webhook, or mailbox. All mainstream competitors (Zapier, Make, Power Automate) charge by volume and require your data to pass through their servers.

Ollama exposes a local LLM on http://localhost:11434 with a compatible HTTP API. Since late 2023, n8n has included a native Ollama node (and an OpenAI-compatible node that works too). Result: a workflow that summarizes client PDFs or sorts HR emails never sends a single byte outside your infrastructure.

#Prerequisites

Ollama installed

Daemon active on port 11434. Check with curl http://localhost:11434. If you’re starting from scratch, first follow the site’s Ollama installation guide.

A downloaded model

ollama pull qwen3.5:9b for versatility, or qwen3.5:4b if you want something lighter. For classification workflows, granite4.2:8b is solid and very token-efficient.

Docker + Docker Compose

n8n deploys in a few seconds with Docker. Docker Desktop on Windows/macOS, Docker Engine on Linux.

8 GB of free RAM

n8n uses ~500 MB, Ollama loads the model into VRAM (~6.6 GB for a 9B Q4). Allow plenty of headroom if you want several workflows running simultaneously.

A little patience for no-code

n8n is visual but not magic: understanding JSON and expressions like {{ $json.field }} remains useful.

#1. Install n8n self-hosted via Docker

The official and simplest method is a Docker image maintained by the n8n team. Create a dedicated directory and a docker-compose.yml file:

services:
  n8n:
    image: docker.n8n.io/n8nio/n8n
    container_name: n8n
    restart: unless-stopped
    ports:
      - "5678:5678"
    environment:
      - N8N_HOST=localhost
      - N8N_PORT=5678
      - N8N_PROTOCOL=http
      - GENERIC_TIMEZONE=Europe/Paris
      - N8N_SECURE_COOKIE=false
    volumes:
      - ./n8n_data:/home/node/.n8n
    extra_hosts:
      - "host.docker.internal:host-gateway"

The extra_hosts line is the key: from the n8n container, you'll call Ollama via http://host.docker.internal:11434 rather than localhost (which would point to the container itself). On Linux only, this trick requires the host-gateway directive.

docker compose up -d
docker compose logs -f n8n

Open http://localhost:5678. On the first connection, n8n asks you to create an owner account—it remains local, with no mandatory telemetry. Once you're logged in, you arrive at the blank canvas.

#2. Configure the Ollama node in n8n

n8n offers two ways to call Ollama: the dedicated “Ollama Chat Model” node (built into n8n’s LangChain branch) and a generic HTTP call. The dedicated node is cleaner for agents and chains; the HTTP call gives you more control. Both work.

1. 01
   Add a Ollama node

   On the canvas, type "/" and search for "Ollama". Choose "Ollama Chat Model" (under AI > Language Models).
2. 02
   Create a credential

   In the Credential field, click "Create new." Base URL: http://host.docker.internal:11434 (from the Docker container) or http://localhost:11434 if n8n runs natively. No API key — it’s local.
3. 03
   Test the connection

   n8n tests automatically and lists the available models. If the list is empty, check with docker exec n8n curl http://host.docker.internal:11434/api/tags from the host.
4. 04
   Choose the model

   In the Model dropdown, select qwen3.5:9b (or the one you pulled). Leave Temperature at the default 0.7; lower it to 0.2 for deterministic tasks (classification, extraction).

#3. First workflow: summarize an RSS feed

Typical case: you follow 15 tech blogs and want a daily summary in French without reading everything. n8n fetches new articles, Ollama summarizes them, and the result is sent by email or written to a Markdown file.

1. 01
   Trigger: Schedule Trigger

   Set it to run every day at 7 a.m. Cron expression mode: 0 7 * * *.
2. 02
   Retrieve the feed: RSS Feed Trigger or RSS Read

   Feed URL (e.g., https://blog.lemondeinformatique.fr/feed/). Enable "Return only new items" to avoid re-summarizing older ones.
3. 03
   Split if needed: Split In Batches

   If the pipeline returns 10 articles, limit it to 5 to avoid saturating Ollama. Batch size: 1, to process articles one at a time.
4. 04
   Call Ollama

   Ollama Chat Model node with this system prompt and the article as a user message.

{
  "system": "Tu es un assistant éditorial. Résume l'article en 3 puces de 15 mots maximum, en français, ton neutre. Pas d'intro, juste les puces.",
  "prompt": "Titre : {{ $json.title }}\n\nContenu : {{ $json.content }}"
}

1. 01
   Aggregate summaries: Merge or Code

   Combine the article outputs into a single email body. A Code node (JavaScript) with items.map(i => `**${i.json.title}**\n${i.json.response}`).join('\n\n') is enough.
2. 02
   Send: Send Email or write file

   Email (SMTP) node to your address, or Write Binary File node to /home/node/.n8n/daily-digest.md (persisted in the Docker volume).

#4. Workflow: translating incoming emails

Typical case: you receive supplier emails in English, German, and Spanish. You want to read them in French without Google Translate (which siphons off the content). n8n monitors your IMAP inbox, Ollama translates, and the result lands in a "Translated" label or as a drafted reply.

1. 01
   Trigger: IMAP Email

   Configure your IMAP credentials (Gmail requires an app password). Filter: all unread emails in the INBOX. Polling every 10 minutes.
2. 02
   Detect language: Ollama #1

   First Ollama call with system prompt: "Respond only with the ISO 639-1 code of the language of this text (fr, en, de, es...). No sentence, just two letters." Prompt: {{ $json.subject }} {{ $json.textPlain }}.
3. 03
   Connect: IF node

   Condition: {{ $json.response }} ≠ "fr". If true → translate. If false → ignore (already in French).
4. 04
   Translate: Ollama #2

   System prompt: "Translate the following text into natural French. Preserve the formatting (paragraphs, lists). Do not comment; translate directly." Prompt: the body of the email.
5. 05
   Final action

   Two options: a Gmail “Add Label” node followed by writing the translation at the top of the message via “Reply Draft,” or simply sending a Slack/Telegram notification with the translated version.

#5. Workflow: classify inbound leads

Typical case: a contact form populates an Airtable / Notion / Postgres database. You want to automatically classify each lead: Hot / Warm / Cold, industry, urgency. n8n listens to the database, Ollama reads the free-form message and writes the structured tags.

This workflow uses Ollama's JSON mode, which forces the output to follow a schema. Essential whenever structured data is involved.

{
  "model": "qwen3.5:9b",
  "format": "json",
  "stream": false,
  "options": { "temperature": 0.1 },
  "system": "Tu es un assistant CRM. Réponds uniquement en JSON valide avec les clés : temperature (Hot/Warm/Cold), secteur (string court), urgence (1-5).",
  "prompt": "Message du lead : {{ $json.message }}\n\nEntreprise : {{ $json.company }}"
}

1. 01
   Trigger: Airtable/Postgres Trigger

   On “New row in table Leads.” Polling every 5 min, or a webhook if your source supports it (more responsive).
2. 02
   Ollama call in JSON mode

   HTTP Request POST node on http://host.docker.internal:11434/api/generate with the body above. The response arrives in $json.response (JSON string).
3. 03
   Parser: Code or Set node

   JSON.parse($input.first().json.response) extracts temperature, sector, and urgency. If parsing fails (because the model hallucinates), retry with a temperature of 0.
4. 04
   Write to the CRM

   Airtable Update / Postgres Update node with the three fields. Optional: if temperature = Hot and urgency ≥ 4, trigger a Slack node that pings commercial@boite.fr.

#Tips & troubleshooting

n8n does not connect to Ollama

ECONNREFUSED on localhost:11434? You are in Docker—use host.docker.internal. On Linux, verify that extra_hosts is present in the compose file and that Ollama is listening on 0.0.0.0 (OLLAMA_HOST=0.0.0.0:11434).

Slow workflow

Ollama loads the model cold (5–15 sec the first time). To keep it warm, run a cron job in parallel that pings /api/generate every 4 minutes, or increase OLLAMA_KEEP_ALIVE=30m.

The model hallucinates during classification

Always use format: "json" plus a system prompt that lists the allowed values. If the 9B in Q4 remains unclear, moving up to qwen3.5:9b-q8_0 significantly improves JSON robustness.

Lost Docker volumes

The workflow and credentials live in ./n8n_data. Never delete this folder without a backup. Export your workflows as JSON via the menu (Settings > Workflows > Download) so you can version them in Git.

Too many calls = VRAM saturation

If a workflow triggers 20 Ollama calls in parallel, the GPU becomes saturated. Use Split In Batches with batch size 1 or add OLLAMA_NUM_PARALLEL=1 to serialize them.

n8n versions

n8n releases a major version every ~10 days. docker compose pull && docker compose up -d updates it. The Ollama node has been stable since 1.18+.

#Go further

You have a working n8n + Ollama pipeline. Some natural next steps:

Connect a RAG

The site’s ChromaDB RAG guide shows how to index your PDFs and notes. You can call this RAG pipeline from n8n through an HTTP endpoint—your workflows become aware of the context in your knowledge base.

Optimize the model

If your workflows process significant volume, read the Q4/Q5/Q8 quantization guide: dropping to Q4_K_M can halve VRAM usage with no visible loss for classification.

Choose a dedicated GPU

If n8n + Ollama need to run 24/7 on one machine, the site's GPU buying guide gives the 12/16/24 GB thresholds for your target workflows.