AI Starter Kit: Ollama + Qdrant + n8n as a Docker Stack

Anyone who wants to build local AI workflows without sending data to cloud providers will find a ready-made Docker Compose template in n8n's Self-hosted AI Starter Kit. It starts n8n as workflow orchestration, Ollama for local language models, Qdrant as a vector database and PostgreSQL as data storage with a single command. In the official n8n documentation the kit is described as a curated collection of AI elements that let you build AI workflows quickly, and the complete setup is available in the GitHub repository n8n-io/self-hosted-ai-starter-kit. There it says the kit is explicitly intended as a starting point and not as a production-ready solution. As of: July 2026.

The four building blocks of the stack

The stack combines four containers, each of which takes on a clearly defined task. According to the repository's README, these are:

• n8n: the low-code platform with over 400 integrations and its own components for AI agents, it orchestrates the workflows and connects the other services with each other.
• Ollama: a cross-platform environment for installing and running current local language models, without requests leaving your own network.
• Qdrant: an open source vector database with a high-performance API that stores embeddings and is used for semantic search as well as for Retrieval-Augmented-Generation workflows.
• PostgreSQL: handles persistent data storage, including for n8n itself and the workflow history.

GPU profiles: Nvidia, AMD or CPU only

The stack can be adapted to the available hardware via Docker Compose profiles. For Nvidia graphics cards, the command docker compose --profile gpu-nvidia up starts the stack with GPU acceleration, on Linux this works the same way for AMD cards via the gpu-amd profile. Without a dedicated GPU, the stack also runs in the CPU profile, in which case the processor handles the model computation, which takes noticeably longer with larger models. On a Mac, the GPU cannot be passed through to Docker directly, here the documentation recommends running Ollama locally on the host system and starting n8n normally via docker compose up.

Installation in a few steps

Getting started follows the usual Docker Compose pattern: clone the repository from GitHub, copy the .env.example file to .env and adjust the secrets and passwords in it, then run the appropriate profile command. After starting, n8n is available at http://localhost:5678/, where a user account is set up once. The kit comes with a bundled example workflow that can be tested directly via the chat button. For data exchange between workflows and the local file system, a shared folder under /data/shared is also mounted into the containers. Anyone keeping the stack up to date uses the corresponding pull and create commands for upgrades before restarting with the respective profile.

What the stack is suited for, and where local models reach their limits

As example workflows, the repository names, among others, AI agents for scheduling, the summarization of PDF documents, the extension of Slack bots as well as the analysis of financial documents, plus templates for tax topics and document evaluation with Qdrant and a Mistral model. The common denominator: sensitive documents such as invoices, contracts or customer data do not leave your own infrastructure. That is the key advantage over cloud APIs. Honestly, local models that run on off-the-shelf hardware are currently still noticeably behind large cloud models in response speed and reasoning quality, especially for complex multi-step tasks. For clearly defined tasks such as classification, summarization or simple extraction, the performance is often sufficient, especially when data protection and control over the infrastructure are more important than maximum model quality. For companies that want to use such an architecture in production and securely, it is usually worth assessing which use cases are suitable for AI agents on their own infrastructure and where a hybrid solution makes more sense. NordFlux supports the selection and setup of such stacks as part of automation projects.

Frequently asked questions about n8n AI Starter Kit, Ollama and Qdrant

Is the Self-hosted AI Starter Kit suitable for production use?

According to the official n8n documentation, explicitly not without further work: The kit is "not fully optimized for production environments" and should be adapted, secured and hardened before productive use.

Do I necessarily need a graphics card to run the stack?

No. There is a dedicated CPU profile that lets the stack run even without a GPU, although model computation then takes longer. Nvidia and AMD graphics cards accelerate Ollama via the respective GPU profiles, on a Mac Ollama is instead run directly on the host.

What license does the Self-hosted AI Starter Kit have?

The repository is under the Apache License 2.0, the full terms can be found in the LICENSE file of the GitHub repository.

Can I integrate my own n8n workflows into the stack?

Yes. The Docker Compose setup serves as a basic framework on which your own workflows, additional integrations and adapted models can be built, the included example workflow only shows the basic structure.