The Master compose deliberately contains no local CPU/GPU NeuroForge workers. It uses remote subagents. Agent, Knowledge and NeuroForge use OLLAMA_BASE_URL/OLLAMA_URLS, normally pointing to the GPU subagent host.

Before first start:

replace the 192.0.2.x example addresses;
fill the five GLPI credentials in .env;
expose the NeuroForge Master port only to trusted subagent networks or put it behind TLS/mTLS/reverse proxy;
keep Grafana and Prometheus bound to loopback unless intentionally proxied.

Start:

./preflight.sh
docker compose --profile research --profile monitoring up -d

Monitoring:

Prometheus: http://127.0.0.1:${PROMETHEUS_HOST_PORT:-9090}
Grafana: http://127.0.0.1:${GRAFANA_HOST_PORT:-3000}
Dashboard is provisioned automatically in folder NeuroForge.