Technical description

Node 20+, no dependencies. The gateway is a micro-router (`node:http`) with SSE; routes are declared as `{ method, path, auth, rate, requiredRole }`. Tenants, keys, budgets and policies live in `config/`, data in `data/`. The model is swappable (openai-compatible provider, fallback chain or mock).

How to run it

node scripts/serve.mjs                            # gateway (NMQ_PORT, default 8787)
NMQ_LLM_PROVIDER=mock node scripts/serve.mjs       # demo without an API key ($0)
node scripts/smoke.mjs                            # 40/42 in this environment (2 known: selfplay, swarm UDP)

curl -X POST http://127.0.0.1:8800/v1/agents/support/run \
  -H "content-type: application/json" -d '{"input":"Kako da resetujem lozinku?"}'

API

Data

• config/tenants.json — tenants, budgets, allowed agents (the demo has one tenant, “demo”)
• config/tools.json — which tools are enabled; the demo disables http_fetch, email_send, notify and all MCP servers
• config/autonomy.json — autonomy levels (demo: L0/L1, no self-scheduling)
• data/ — sessions, history, vector store, audit, cost (per tenant)

Limits — what this demo does not do

• The demo uses a mock model: the text is not from a real model, the machinery is.
• The demo disables http_fetch, email_send, notify, invoice_create and MCP; admin, keys, chaos, worker and audit routes are blocked at nginx.
• One demo tenant, a small budget, a 20/min rate limit — the demo is not for load.
• Without a key there is no live model; mock is deliberate so the demo cannot spend money.

How it is deployed

One Node process behind nginx, a systemd unit, Let’s Encrypt, demo configuration from `demo-config/`. Recipe: E:\poslovi\DEMO-REGISTAR.md.