Remote OpenClaw Blog

Running OpenClaw 100% Locally: The Complete Self-Hosted AI Agent Stack

9 min read · 7 April 2026

Why Go Fully Local?

I run my primary OpenClaw instance on a Mac Mini M4 sitting on a shelf in my home office. No AWS bills. No OpenAI API keys. No data leaving my local network. After six months of running this setup, I can tell you the three reasons that actually matter:

Privacy. When OpenClaw processes my emails, calendar events, and client documents, none of that data hits a third-party server. For anyone handling client data, financial information, or anything covered by an NDA, this is not a nice-to-have — it is a requirement. I work with consulting clients who would not be comfortable knowing their project details are being sent to OpenAI or Anthropic's servers for processing, even with those companies' privacy policies.

Cost predictability. Cloud API costs scale with usage, and they can surprise you. A local setup has a fixed upfront cost (the hardware) and near-zero ongoing cost (electricity). After about 4-6 months of moderate use, the Mac Mini pays for itself compared to cloud API costs. I did the math and share the full breakdown later in this guide.

Control. No rate limits. No API outages at 2 AM when your scheduled skills are supposed to run. No model deprecation notices forcing you to migrate. The model you downloaded today works the same way in six months. As the Contabo blog and Emergent.sh have both covered, self-hosting AI infrastructure is becoming increasingly viable for individuals and small teams.

Hardware Requirements

Here is what I recommend based on my own testing and what I have seen work for other operators in the community:

I use the Mac Mini M4 with 24GB. The Apple Silicon unified memory architecture is uniquely good for LLM inference because the CPU and GPU share the same memory pool — no bottleneck copying data between CPU RAM and GPU VRAM. This is why a $799 Mac Mini can match or beat a $1,500 Linux setup with a dedicated GPU for most LLM workloads.

For a detailed Mac Mini setup walkthrough, see the OpenClaw Mac Mini setup guide.

Ollama Setup

Ollama is the runtime that serves local LLMs. It handles model downloading, quantization, and provides an OpenAI-compatible API that OpenClaw connects to.

Install Ollama

# macOS
brew install ollama

# Linux
curl -fsSL https://ollama.ai/install.sh | sh

# Start the Ollama server
ollama serve

Pull Your Models

# Primary reasoning model
ollama pull qwen3:8b

# Fast model for simple tasks
ollama pull glm4:9b

# Embedding model for vector search
ollama pull nomic-embed-text

# Verify all models are available
ollama list

Configure OpenClaw to Use Ollama

# ~/.openclaw/config.yaml
llm:
  provider: ollama
  model: qwen3:8b
  base_url: http://localhost:11434
  temperature: 0.7
  max_tokens: 4096

embedding:
  provider: ollama
  model: nomic-embed-text
  base_url: http://localhost:11434

That is it. No API keys, no account creation, no billing setup. OpenClaw connects to the local Ollama server and uses whatever models you have pulled.

Recommended Local Models

Not all local models are equal for agent tasks. After testing over a dozen models on my Mac Mini, here are the two I recommend:

Qwen3 8B — Best All-Around

Qwen3 8B is my daily driver for OpenClaw. It handles instruction-following, code generation, summarization, and structured output extraction with quality that genuinely surprised me. On my Mac Mini M4 (24GB), it generates about 40 tokens per second — fast enough that scheduled tasks complete quickly and interactive queries feel responsive.

The key advantage of Qwen3 8B over other 8B models is its instruction-following precision. When an OpenClaw skill asks it to extract specific fields from an email and return them as JSON, Qwen3 8B does it correctly about 95% of the time. Smaller or less capable models often hallucinate extra fields, miss required ones, or return malformed JSON. See the Qwen3 8B OpenClaw guide for benchmarks and tuning tips.

GLM-4.7-flash — Fastest Inference

GLM-4.7-flash is the speed champion. It generates 55-65 tokens per second on my hardware — roughly 50% faster than Qwen3 8B. The quality is slightly lower on complex reasoning tasks, but for high-frequency, low-complexity tasks (email triage, data extraction, simple summaries), the speed advantage makes it the better choice.

I use GLM-4.7-flash for my email scanning and calendar sync skills (which run many times per day) and Qwen3 8B for weekly reports and complex analysis skills. OpenClaw lets you assign different models to different skills, so you can optimize each workflow independently.

Docker Compose Configuration

If you prefer a containerized setup (especially useful on Linux), here is a Docker Compose configuration that runs the entire stack:

# docker-compose.yaml
version: '3.8'

services:
  ollama:
    image: ollama/ollama:latest
    container_name: ollama
    ports:
      - "11434:11434"
    volumes:
      - ollama_data:/root/.ollama
    restart: unless-stopped
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]  # Remove this block for CPU-only

  openclaw:
    image: openclaw/openclaw:latest
    container_name: openclaw
    depends_on:
      - ollama
    environment:
      - OLLAMA_BASE_URL=http://ollama:11434
      - OPENCLAW_DATA_DIR=/data
    volumes:
      - openclaw_data:/data
      - ./config:/config
    restart: unless-stopped

volumes:
  ollama_data:
  openclaw_data:

Starting the Stack

# Start everything
docker compose up -d

# Pull models inside the Ollama container
docker exec ollama ollama pull qwen3:8b
docker exec ollama ollama pull nomic-embed-text

# Verify OpenClaw is connected
docker exec openclaw openclaw status

The Docker setup is particularly clean on Linux servers. On macOS, I prefer running Ollama natively (via Homebrew) and OpenClaw directly, since Docker on macOS adds a virtualization layer that slightly reduces inference performance.

# Install Tailscale on your Mac Mini
brew install tailscale

# Start Tailscale and authenticate
tailscale up

# Your Mac Mini gets an IP like 100.x.y.z
tailscale ip

# From your laptop or phone (on Tailscale)
curl http://100.x.y.z:11434/api/tags  # Check Ollama models
curl http://100.x.y.z:8080/status     # Check OpenClaw status