Local AI: Hermes Agent on the Mini PC

Get the eGPU ready

Keep the eGPU always connected and install a current NVIDIA driver. Then check it in PowerShell:

nvidia-smi

If it lists the RTX 4070 Ti, the driver is working.

Install Hermes in a container

Hermes will be always on, so it runs fully walled off from Ben's Windows files.

1. Install Docker Desktop with the WSL2 backend.
2. Save this as docker-compose.yml:

services:
  hermes:
    image: nousresearch/hermes-agent:stable
    container_name: hermes
    restart: unless-stopped
    command: gateway run
    ports:
      - "127.0.0.1:9119:9119"
    volumes:
      - hermes-data:/opt/data
    environment:
      - HERMES_DASHBOARD=1
      - HERMES_DASHBOARD_BASIC_AUTH_USERNAME=ben
      - HERMES_DASHBOARD_BASIC_AUTH_PASSWORD=${HERMES_PW}
      - HERMES_DASHBOARD_BASIC_AUTH_SECRET=${HERMES_SECRET}
    shm_size: 1g
    deploy:
      resources:
        limits:
          memory: 4G
          cpus: "2.0"
volumes:
  hermes-data:

Put HERMES_PW and HERMES_SECRET in a .env file next to the compose file.

3. Run setup and pick Full Setup, then start it:

docker compose run --rm hermes setup
docker compose up -d

4. Open Hermes Desktop and connect it as a Remote Gateway to http://localhost:9119. Sign in with the basic-auth user.

Data lives in the Docker volume hermes-data, not a Windows folder. That also avoids SQLite corruption on Windows bind mounts.

Local models: run Ollama on Windows so it uses the 4070 Ti directly. In setup, choose Custom endpoint with base_url http://host.docker.internal:11434/v1 and at least 64K context.

Simpler but not isolated: the native Hermes Desktop installer from hermes-agent.nousresearch.com.

Import OpenClaw inside the container

Setup already ran in step 2. The container can't see Windows, so first copy OpenClaw's folder in. Only that folder is exposed, nothing else from Windows.

docker cp <path-to-openclaw-folder> hermes:/opt/data/import/

Then install the importers and do a dry run:

docker exec -it hermes hermes plugins install agent-importers
docker exec -it hermes hermes import-agent openclaw --dry-run

If it looks right, run it for real. --migrate-secrets brings the API keys across.

docker exec -it hermes hermes import-agent openclaw --migrate-secrets

Add a local model

• 64K min context
• 7–14B at Q4
• 12 GB VRAM

Sweet spot for 12 GB: 7–14B models at Q4, fully on the GPU.

Ollama runs on Windows, not in the container. Pull a model there, then pick it in Hermes Desktop (Settings > Providers) or with docker exec -it hermes hermes model. Pick the star or green rows.

Set at least 64K tokens of context:

--ctx-size 65536   # llama.cpp
-c 65536           # Ollama

Skip 27–35B models: they won't fit in 12 GB, spill into system RAM and run slowly. See model sizing.

Connect Home Assistant

docker exec -it hermes hermes plugins install homeassistant

Create a long-lived access token in Home Assistant (profile > Security). Put HASS_TOKEN and HASS_URL in /opt/data/.env, and set watch_domains in /opt/data/config.yaml.

HASS_URL must be reachable from the container: http://homeassistant.local:8123 or the HA IP, or http://host.docker.internal:8123 if HA runs on the same PC.

Then restart. The gateway already runs, so there's no separate hermes gateway.

docker compose restart hermes

Optional extras and checks

Mostly done by step 2: Hermes already runs isolated in its container. Optional extras:

Egress proxy, to control what Hermes can reach online:

docker exec -it hermes hermes egress setup
docker exec -it hermes hermes egress start

Check the setup:

docker exec -it hermes hermes doctor