{"id":3603,"date":"2026-04-03T23:20:45","date_gmt":"2026-04-03T23:20:45","guid":{"rendered":"https:\/\/www.tooljunction.io\/blog\/?p=3603"},"modified":"2026-04-03T23:20:47","modified_gmt":"2026-04-03T23:20:47","slug":"self-hosted-ai-stack-2026","status":"publish","type":"post","link":"https:\/\/www.tooljunction.io\/blog\/self-hosted-ai-stack-2026","title":{"rendered":"How to Build Your Self Hosted AI Stack in 2026 (Ollama + Open WebUI + n8n)"},"content":{"rendered":"\n

You don’t need a $200\/month cloud AI subscription. You need a $7 VPS and 30 minutes.<\/strong><\/p>\n\n\n\n

The self-hosted AI movement has gone mainstream. What used to require a GPU cluster and a PhD now runs on a single VPS with Docker Compose. In 2026, you can run your own ChatGPT-like interface, serve local LLMs, and automate AI workflows without sending a single byte of data to anyone else’s servers.<\/p>\n\n\n\n

This guide walks you through the exact stack: Ollama<\/a><\/strong> for running models locally, Open WebUI<\/a><\/strong> for the chat interface, and n8n<\/a><\/strong> for AI-powered workflow automation. All self hosted. All free. All yours.<\/p>\n\n\n\n

Why Self-Host Your AI Tools?<\/h2>\n\n\n\n

Every time you send a prompt to ChatGPT, Claude, or Gemini, your data hits someone else’s server. For personal projects, that’s fine. For business data, client information, or proprietary code? That’s a risk most people don’t think about until it’s too late.<\/p>\n\n\n\n

Self-hosting your AI stack gives you three things that cloud services can’t:<\/p>\n\n\n\n

Data sovereignty.<\/strong> Your prompts, documents, and conversations never leave your network. No third-party logging, no training on your data, no surprises in the terms of service.<\/p>\n\n\n\n

Zero recurring costs.<\/strong> Once your server is running, you’re done. No per-token pricing, no execution limits, no “you’ve hit your free tier” popups. Run as many queries as your hardware can handle.<\/p>\n\n\n\n

Full control.<\/strong> Pick your models. Tune your parameters. Build custom workflows. Swap components whenever something better shows up. No vendor lock-in, no feature gates, no waiting for someone else’s roadmap.<\/p>\n\n\n\n

The Stack: What Each Tool Does<\/h2>\n\n\n\n

Before we get into setup, here’s how the three pieces fit together:<\/p>\n\n\n\n

Tool<\/th>Role<\/th>Replaces<\/th><\/tr><\/thead>
Ollama<\/strong><\/td>Local LLM inference server<\/td>OpenAI API, Anthropic API<\/td><\/tr>
Open WebUI<\/strong><\/td>Chat interface with RAG, voice, and vision<\/td>ChatGPT, Perplexity<\/td><\/tr>
n8n<\/strong><\/td>AI workflow automation<\/td>Zapier + AI plugins, Make<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

Ollama is the engine. It downloads, manages, and serves open-weight language models through a simple API. Open WebUI is the dashboard. It gives you a polished ChatGPT-style interface that connects to Ollama (or any OpenAI-compatible API). n8n is the glue. It connects your AI models to 400+ apps and services through visual workflows.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Together, they form a complete self hosted AI platform that rivals what you’d get from paying $200+\/month across multiple SaaS subscriptions.<\/p>\n\n\n\n

Step 1: Choose Your Hardware<\/h2>\n\n\n\n

You don’t need a monster rig. Here’s what actually works:<\/p>\n\n\n\n

Budget option (good for small models):<\/strong> Any VPS with 4+ GB RAM and 2 vCPUs. Providers like Hetzner<\/a>, DigitalOcean, or Contabo offer this for $5 to $10\/month. You’ll run smaller models like Gemma 3 1B or Phi-4 Mini comfortably.<\/p>\n\n\n\n

Recommended setup (covers most use cases):<\/strong> A VPS or home server with 8 to 16 GB RAM and 4 vCPUs. This handles 7B to 8B parameter models like Llama 3.3 8B and Qwen 2.5 7B at usable speeds. If you’re already running a Hetzner box for other projects, this is the sweet spot.<\/p>\n\n\n\n

Power user (GPU inference):<\/strong> A dedicated GPU server or a local machine with an NVIDIA RTX 3060+ (12 GB VRAM) or Apple Silicon with 16 GB+ unified memory. This unlocks 14B to 27B parameter models that rival GPT-4 quality on many tasks.<\/p>\n\n\n\n

For most indie developers and small teams, the 8 GB VPS option is the right starting point. You can always scale up later.<\/p>\n\n\n\n

Step 2: Install Ollama<\/h2>\n\n\n\n

Ollama is a one-liner install on Linux:<\/p>\n\n\n\n

curl -fsSL https:\/\/ollama.com\/install.sh | sh\n<\/code><\/pre>\n\n\n\n
Verify it’s running:<\/p>\n\n\n\n
ollama --version\n<\/code><\/pre>\n\n\n\n
Now pull your first model. For a good balance of quality and speed on modest hardware:<\/p>\n\n\n\n
# Lightweight, runs on 4 GB RAM\nollama pull gemma3:1b\n\n# Best all-rounder for 8 GB+ RAM\nollama pull llama3.3:8b\n\n# Strong reasoning and coding\nollama pull qwen2.5:7b\n\n# If you have 16 GB+ RAM and want near-GPT-4 quality\nollama pull qwen2.5:14b\n<\/code><\/pre>\n\n\n\n
Ollama serves an API on http:\/\/localhost:11434<\/code> by default. That’s all Open WebUI and n8n need to connect.<\/p>\n\n\n\n
Pro tip:<\/strong> If you’re running Ollama on a VPS and want other services in Docker to access it, you’ll need to set OLLAMA_HOST=0.0.0.0<\/code> in your environment. But never expose port 11434 to the public internet without authentication. Keep it behind your firewall or reverse proxy.<\/p>\n\n\n\n
Step 3: Deploy Open WebUI<\/h2>\n\n\n\n
Open WebUI turns Ollama into a full ChatGPT alternative with a clean web interface. One Docker command gets you running:<\/p>\n\n\n\n
docker run -d \\\n  -p 3000:8080 \\\n  --add-host=host.docker.internal:host-gateway \\\n  -v open-webui:\/app\/backend\/data \\\n  --name open-webui \\\n  --restart always \\\n  ghcr.io\/open-webui\/open-webui:main\n<\/code><\/pre>\n\n\n\n
Open http:\/\/your-server-ip:3000<\/code> in your browser. Create an admin account on first launch.<\/p>\n\n\n\n
What You Get Out of the Box<\/h3>\n\n\n\n
Open WebUI has grown into a serious platform. It’s not just a chat wrapper anymore. Here’s what ships by default:<\/p>\n\n\n\n
    \n
  1. RAG (Retrieval-Augmented Generation).<\/strong> Upload PDFs, Word docs, or text files and chat with them directly. Open WebUI builds a local knowledge base from your documents, so your models can answer questions grounded in real data instead of hallucinating.<\/li>\n\n\n\n
  2. Web search integration.<\/strong> Connect a search engine (SearXNG<\/a> is the self-hosted favorite) and your local models can pull live information from the web. This bridges the biggest gap between local and cloud AI.<\/li>\n\n\n\n
  3. Voice input and output.<\/strong> Built-in speech-to-text (via Whisper) and text-to-speech support. You can talk to your self-hosted AI just like you would with ChatGPT’s voice mode.<\/li>\n\n\n\n
  4. Multi-model support.<\/strong> Run Llama, Qwen, Gemma, DeepSeek, and Mistral side by side. Compare outputs in the same window. Switch between models mid-conversation.<\/li>\n\n\n\n
  5. Code execution.<\/strong> A built-in Python interpreter (pyodide) lets your AI write and run code right in the chat. Useful for data analysis, quick scripts, and prototyping.<\/li>\n\n\n\n
  6. User management.<\/strong> RBAC (role-based access control), SSO support, and audit logs. This matters when you’re running it for a team, not just yourself.<\/li>\n<\/ol>\n\n\n\n
    Connecting Open WebUI to Cloud APIs (Optional)<\/h3>\n\n\n\n
    The beauty of Open WebUI is that it’s provider-agnostic. You can connect it to Ollama for local models AND to OpenAI, Anthropic, or any OpenAI-compatible API at the same time. This gives you a hybrid setup: use local models for everyday tasks (free, private) and cloud models for the occasional heavy-lifting query.<\/p>\n\n\n\n
    Go to Settings > Connections in Open WebUI to add your API keys.<\/p>\n\n\n\n
    Step 4: Set Up n8n for AI Workflow Automation<\/h2>\n\n\n\n
    This is where the stack goes from “cool personal tool” to “actual productivity multiplier.”<\/p>\n\n\n\n
    n8n is a self-hosted workflow automation platform. Think Zapier, but you own it, there are no per-execution fees, and it has native AI nodes that connect directly to your Ollama models.<\/p>\n\n\n\n
    Deploy n8n with Docker Compose<\/h3>\n\n\n\n
    Create a docker-compose.yml<\/code>:<\/p>\n\n\n\n
    version: '3.8'\nservices:\n  n8n:\n    image: n8nio\/n8n:latest\n    restart: always\n    ports:\n      - \"5678:5678\"\n    environment:\n      - N8N_HOST=localhost\n      - N8N_PORT=5678\n      - N8N_PROTOCOL=http\n    volumes:\n      - n8n_data:\/home\/node\/.local\/share\/n8n\n    extra_hosts:\n      - \"host.docker.internal:host-gateway\"\n\nvolumes:\n  n8n_data:\n<\/code><\/pre>\n\n\n\n
    docker compose up -d\n<\/code><\/pre>\n\n\n\n
    Now, open http:\/\/your-server-ip:5678<\/code> and create your account.<\/p>\n\n\n\n
    Connect n8n to Ollama<\/h3>\n\n\n\n
    In n8n, add Ollama credentials:<\/p>\n\n\n\n
      \n
    1. Go to Credentials > New Credential > Ollama<\/li>\n\n\n\n
    2. Set the base URL to http:\/\/host.docker.internal:11434<\/code><\/li>\n\n\n\n
    3. Select your model (e.g., llama3.3)<\/li>\n\n\n\n
    4. Save and test the connection<\/li>\n<\/ol>\n\n\n\n
      AI Workflow Ideas That Actually Save Time<\/h3>\n\n\n\n
      Here are real workflows you can build in 30 minutes or less:<\/p>\n\n\n\n
        \n
      1. Email triage.<\/strong> Trigger on new email > send content to Ollama > classify as urgent\/normal\/spam > route to the right folder or send a Slack notification. Saves 20+ minutes of inbox sorting every day.<\/li>\n\n\n\n
      2. Content summarizer.<\/strong> Watch an RSS feed or Slack channel > extract new posts > summarize with your local LLM > post the summary to a Notion database or Discord channel. Great for staying on top of industry news without reading everything.<\/li>\n\n\n\n
      3. Document Q&A bot.<\/strong> Upload company docs to Open WebUI’s knowledge base > create an n8n webhook that receives questions via API > query the knowledge base > return answers. Your own internal ChatGPT trained on your docs.<\/li>\n\n\n\n
      4. Code review assistant.<\/strong> Trigger on new GitHub PR > fetch the diff > send to Ollama for review > post comments on the PR. Works surprisingly well with Qwen 2.5 Coder models.<\/li>\n\n\n\n
      5. Lead qualification.<\/strong> Webhook receives form submission > Ollama analyzes the message > scores and categorizes the lead > creates a task in your project management tool. No Zapier subscription required.<\/li>\n<\/ol>\n\n\n\n
        The Complete Docker Compose Stack<\/h2>\n\n\n\n
        Want everything in one file? Here’s the full stack:<\/p>\n\n\n\n
        version: '3.8'\n\nservices:\n  ollama:\n    image: ollama\/ollama:latest\n    restart: always\n    ports:\n      - \"11434:11434\"\n    volumes:\n      - ollama_data:\/root\/.ollama\n    # Uncomment for NVIDIA GPU support:\n    # deploy:\n    #   resources:\n    #     reservations:\n    #       devices:\n    #         - driver: nvidia\n    #           count: all\n    #           capabilities: [gpu]\n\n  open-webui:\n    image: ghcr.io\/open-webui\/open-webui:main\n    restart: always\n    ports:\n      - \"3000:8080\"\n    environment:\n      - OLLAMA_BASE_URL=http:\/\/ollama:11434\n    volumes:\n      - open_webui_data:\/app\/backend\/data\n    depends_on:\n      - ollama\n\n  n8n:\n    image: n8nio\/n8n:latest\n    restart: always\n    ports:\n      - \"5678:5678\"\n    environment:\n      - N8N_HOST=localhost\n      - N8N_PORT=5678\n      - N8N_PROTOCOL=http\n    volumes:\n      - n8n_data:\/home\/node\/.local\/share\/n8n\n    depends_on:\n      - ollama\n\nvolumes:\n  ollama_data:\n  open_webui_data:\n  n8n_data:\n<\/code><\/pre>\n\n\n\n
        Save this as docker-compose.yml<\/strong><\/code>, run docker compose up -d<\/strong><\/code>, and your entire self-hosted AI stack is live.<\/p>\n\n\n\n
        After startup, pull a model into Ollama:<\/p>\n\n\n\n
        \ndocker exec -it ollama ollama pull llama3.3:8b<\/code><\/pre>\n\n\n\n
        <\/p>\n\n\n\n
        \"\"<\/figure>\n\n\n\n
          Example of AI Email Triage workflow<\/em><\/p>\n\n\n\n
        Which Models Should You Run?<\/h2>\n\n\n\n
        The model landscape moves fast. Here’s a practical cheat sheet for April 2026:<\/p>\n\n\n\n
        Model<\/th>Size<\/th>RAM Needed<\/th>Best For<\/th><\/tr><\/thead>
        Gemma 3 1B<\/td>1.6 GB<\/td>4 GB<\/td>Quick tasks, edge devices<\/td><\/tr>
        Llama 3.3 8B<\/td>4.7 GB<\/td>8 GB<\/td>General purpose, good all-rounder<\/td><\/tr>
        Qwen 2.5 7B<\/td>4.4 GB<\/td>8 GB<\/td>Multilingual, strong reasoning<\/td><\/tr>
        Phi-4 14B<\/td>9 GB<\/td>16 GB<\/td>Math, logic, structured reasoning<\/td><\/tr>
        Qwen 2.5 Coder 14B<\/td>9 GB<\/td>16 GB<\/td>Code generation, code review<\/td><\/tr>
        DeepSeek R1<\/td>varies<\/td>16 GB+<\/td>Deep reasoning, chain-of-thought<\/td><\/tr>
        Qwen 3.5 27B<\/td>17 GB<\/td>24 GB+<\/td>Near-GPT-4 quality, the current king<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
        Start with Llama 3.3 8B. It’s the default for a reason: fast, capable, and runs on almost anything. Upgrade to Qwen 2.5 14B or Qwen 3.5 27B when you need better output quality and have the RAM to support it.<\/p>\n\n\n\n
        Securing Your Stack<\/h2>\n\n\n\n
        Running AI tools on a public server means you need basic security hygiene:<\/p>\n\n\n\n
        Use a reverse proxy.<\/strong> Put Nginx or Traefik in front of everything. Get free SSL certificates from Let’s Encrypt. Never expose raw ports to the internet.<\/p>\n\n\n\n
        Lock down Ollama.<\/strong> By default, Ollama has no authentication. Keep it bound to localhost or your Docker network. Don’t expose port 11434 publicly.<\/p>\n\n\n\n
        Enable authentication everywhere.<\/strong> Open WebUI has built-in user management. n8n has its own auth. Use strong passwords and enable 2FA where available.<\/p>\n\n\n\n
        Keep things updated.<\/strong> All three projects ship frequent updates. Set a reminder to pull new images monthly:<\/p>\n\n\n\n
        docker compose pull\ndocker compose up -d\n<\/code><\/pre>\n\n\n\n
        Firewall rules.<\/strong> Only expose ports 80\/443 (via your reverse proxy) to the world. Everything else stays internal.<\/p>\n\n\n\n
        What This Stack Costs<\/h2>\n\n\n\n
        Let’s compare:<\/p>\n\n\n\n
        Setup<\/th>Monthly Cost<\/th><\/tr><\/thead>
        ChatGPT Pro<\/a> + Zapier Pro<\/a> + API usage<\/td>$90 to $200+<\/td><\/tr>
        Self-hosted (Hetzner CPX21, 4 GB)<\/td>~$7<\/td><\/tr>
        Self-hosted (Hetzner CPX31, 8 GB)<\/td>~$13<\/td><\/tr>
        Self-hosted (home server, already owned)<\/td>$0 (electricity only)<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
        The self-hosted stack eliminates per-token pricing, per-execution fees, and monthly subscription costs. Your only recurring expense is the server itself.<\/p>\n\n\n\n
        When Self-Hosting Isn’t the Right Call<\/h2>\n\n\n\n
        Let’s be honest about the tradeoffs:<\/p>\n\n\n\n
        You need frontier-model quality for everything.<\/strong> GPT-4o, Claude Opus, and Gemini Ultra are still ahead of local models on the hardest tasks. If your work requires consistently top-tier reasoning across complex, multi-step problems, cloud APIs are still worth it for those specific queries.<\/p>\n\n\n\n
        You don’t want to maintain infrastructure.<\/strong> Self-hosting means you’re the sysadmin. Updates, backups, monitoring, and debugging are your responsibility. If that sounds like a chore rather than a feature, managed solutions exist for every tool in this stack.<\/p>\n\n\n\n
        You need real-time collaboration at scale.<\/strong> For large teams (50+ people) with complex access control requirements, enterprise SaaS products have battle-tested features that take significant effort to replicate with self-hosted tools.<\/p>\n\n\n\n
        The sweet spot? Run self-hosted for 90% of your AI usage (drafting, summarizing, coding, automating) and keep a cloud API key for the 10% that actually needs frontier models.<\/p>\n\n\n\n
        What to Do Next<\/h2>\n\n\n\n
          \n
        1. Spin up the stack.<\/strong> Copy the Docker Compose file above and deploy it on a $7 VPS. The whole process takes under 30 minutes.<\/li>\n\n\n\n
        2. Pull a model.<\/strong> Start with Llama 3.3 8B. Run a few conversations in Open WebUI to get a feel for local AI quality.<\/li>\n\n\n\n
        3. Build one n8n workflow.<\/strong> Pick the most repetitive task in your day and automate it. Email triage is a great first win.<\/li>\n\n\n\n
        4. Explore the ecosystem.<\/strong> <\/strong>The selfh.st\/apps<\/a> directory tracks hundreds of self-hosted tools. Browse it for companion apps that extend your stack: SearXNG for private search, Paperless-ngx for document management, Immich for photo backup.<\/strong><\/li>\n<\/ol>\n\n\n\n
          The self-hosted AI ecosystem in 2026 is mature, accessible, and genuinely useful. You don’t need to be a sysadmin to run it. You just need Docker, a cheap server, and 30 minutes.<\/p>\n\n\n\n
          Your data. Your models. Your rules.<\/p>\n\n\n\n
          \n\n\n\n
          Looking for more self-hosted tools and AI resources? Check out ToolJunction<\/a> for a curated directory of the best AI tools and open-source software.<\/em><\/p>\n\n\n\n
          <\/p>\n","protected":false},"excerpt":{"rendered":"
          You don’t need a $200\/month cloud AI subscription. This guide walks you through building a complete self-hosted AI stack with Ollama for local LLMs, Open WebUI for a ChatGPT-like interface, and n8n for workflow automation. All running on a single VPS with Docker Compose<\/p>\n","protected":false},"author":1,"featured_media":3604,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7],"tags":[],"class_list":["post-3603","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-guide"],"_links":{"self":[{"href":"https:\/\/www.tooljunction.io\/blog\/wp-json\/wp\/v2\/posts\/3603","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.tooljunction.io\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.tooljunction.io\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.tooljunction.io\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.tooljunction.io\/blog\/wp-json\/wp\/v2\/comments?post=3603"}],"version-history":[{"count":1,"href":"https:\/\/www.tooljunction.io\/blog\/wp-json\/wp\/v2\/posts\/3603\/revisions"}],"predecessor-version":[{"id":3607,"href":"https:\/\/www.tooljunction.io\/blog\/wp-json\/wp\/v2\/posts\/3603\/revisions\/3607"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.tooljunction.io\/blog\/wp-json\/wp\/v2\/media\/3604"}],"wp:attachment":[{"href":"https:\/\/www.tooljunction.io\/blog\/wp-json\/wp\/v2\/media?parent=3603"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.tooljunction.io\/blog\/wp-json\/wp\/v2\/categories?post=3603"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.tooljunction.io\/blog\/wp-json\/wp\/v2\/tags?post=3603"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}