A multi-tenant AI agent that acts, not just answers
NMQ Robot plugs into any site as a JS widget or into an internal process as a REST/SSE API. Its capabilities come from tools, memory and orchestration — and every action passes through policy, budget and an audit trail. In the demo the model is replaced by a “mock” provider: you see the flow, the routing, the approvals, the cost and the audit, but not the text quality of a real model.
The problem it solves
- An agent that answers nicely but cannot do anything — no tools, no memory, no trace of what it did.
- Once it is in production nobody knows what it costs, who approved a risky action, and what exactly ran.
- One client, one agent, one key — while you need many without mixing their data.
What you get
Widget and API at once
One script tag on a site and the same agent over REST/SSE for internal processes. Tenants, keys and rate limits are built in.
6 orchestration patterns
Router, sequential, orchestrator–worker, fanout, handoff and magentic — plus a critic that reviews the answer. Agents are defined by data (JSON), not by code.
Supervision you can prove
Policies (allow / deny / require approval), budget per tenant and per run, PII redaction, traces/spans, metrics and a hash-chained audit.
Memory per tenant
Session, history and a vector store separated per tenant; the built-in offline embedder works without a key and without cost.