n8n Self-Hosted AI Starter Kit Explained

What the Starter Kit Includes

The n8n Self-Hosted AI Starter Kit is an open-source Docker Compose template maintained by the n8n team that bundles everything you need to run AI workflows locally. The kit includes four core services: n8n itself for workflow orchestration, Ollama for running local LLMs, Qdrant as a vector database for semantic search and RAG, and PostgreSQL for relational data storage and conversation memory. All four services are pre-configured to communicate with each other, so you get a working AI automation stack with a single docker compose up command.

The kit is licensed under Apache 2.0, which is separate from n8n's fair-code license. This means there are no commercial restrictions on how you deploy or use the starter kit. You can build commercial products on top of it, modify it freely, and redistribute your changes.

The repository includes pre-built workflow templates that demonstrate common AI patterns. These templates cover conversational AI agents with memory, RAG pipelines for document Q&A, and multi-tool agents that combine web search, database queries, and LLM reasoning. You can import these templates directly into n8n and modify them for your specific use case rather than building from scratch.

Hardware Requirements and Profiles

The starter kit supports multiple hardware configurations through Docker Compose profiles. The default profile runs everything on CPU, which works for testing and lightweight workloads but produces noticeably slow inference with larger models. The gpu-nvidia profile enables CUDA acceleration for NVIDIA GPUs, which is the recommended setup for any production use. An AMD GPU profile is also available for users with Radeon hardware.

Minimum hardware for the CPU profile is 8GB of RAM, though 16GB is recommended. For GPU-accelerated setups, you need an NVIDIA GPU with at least 8GB of VRAM to run 7B parameter models comfortably, or 16GB+ VRAM for 13B parameter models. Mac users can run Ollama natively (outside Docker) for Metal GPU acceleration while connecting it to the dockerized n8n, Qdrant, and PostgreSQL services.

Storage requirements depend on the models you download. Each 7B parameter model typically requires 4 to 5GB of disk space. Qdrant vector storage grows based on the number of documents you ingest. A reasonable starting point is 50GB of free disk space, which accommodates several models and a moderate document collection.

Setting Up the Starter Kit

Prerequisites are Docker and Docker Compose installed on your machine. Clone the repository from GitHub, copy the example environment file to .env, and run docker compose up with the appropriate profile flag for your hardware. The first run will download the Docker images and may take several minutes depending on your internet connection.

Once running, n8n is accessible at localhost:5678. Ollama runs on port 11434, Qdrant on port 6333, and PostgreSQL on port 5432. The n8n instance comes pre-configured with credential entries for Ollama and Qdrant, so you can start building AI workflows immediately without manual connection setup.

You will need to pull at least one model into Ollama before running AI workflows. Common choices are llama3.1 for general reasoning, nomic-embed-text for generating embeddings in RAG pipelines, and mistral for a good balance of speed and quality. The kit documentation includes recommended model combinations for different use cases.

Community Extensions

The most popular community extension is the fork by developer Cole Medin, which significantly expands the starter kit. His version adds Supabase for structured data storage with built-in auth, Open WebUI as a ChatGPT-style frontend for interacting with local models, Flowise as an alternative visual AI builder, Neo4j for graph database support (useful for knowledge graphs), Langfuse for LLM observability and cost tracking, SearXNG for privacy-respecting web search, and Caddy as a reverse proxy for HTTPS access.

The extended kit also includes pre-built agent workflow templates that go beyond the basics. These templates implement patterns like research agents that combine web search with RAG, coding assistants that can read and write files, and multi-step analysis pipelines that process documents through sequential AI stages.

Other community contributions include Helm charts for Kubernetes deployment, Terraform modules for cloud provisioning, and monitoring dashboards using Grafana and Prometheus. The ecosystem around the starter kit has grown substantially since its release, reflecting the community demand for self-hosted AI infrastructure.

What the Kit Does Not Cover

The starter kit is designed for development and small-scale production. It does not include load balancing, horizontal scaling, or high-availability configurations. For production deployments serving many concurrent users, you will need to add a reverse proxy (Caddy or nginx), configure PostgreSQL replication, and potentially run multiple n8n worker instances.

Authentication is minimal in the default configuration. n8n has its own user management, but the other services (Ollama, Qdrant) are exposed without authentication. In a production environment, you should restrict network access to these services or add authentication layers.

Model management is manual. You need to pull models into Ollama yourself and manage disk space as your model collection grows. There is no built-in model registry, version management, or automatic cleanup of unused models.