Ask MIRA Anything. 36 Specialists Answer.

One. MIRA is a single plant intelligence, not 36 separate bots. A supervisor reads each question and routes it to the right specialist, and the agents share one set of plant records, one permission model and one Decision Trail. Your team has one conversation with specialist depth, and every answer links to the agent run that produced it.

MIRA uses a human-in-the-loop system. Each action type has a confidence threshold that your team sets. Below it, or when confidence is unknown, MIRA presents its recommendation, confidence and evidence, and waits for a named approver. A question alone never triggers an action, and MIRA never silently executes a risky one.

Yes. MIRA supports OpenAI, Azure OpenAI, Anthropic, Google, Mistral and any OpenAI-compatible endpoint, including a model hosted on your own servers. Each agent can use a different model, so cost and capability can be matched per task, and the cost of every agent run is measured and shown.

Yes. The deterministic analytics all still run: equipment health, maintenance KPIs, spares, Weibull, OEE, energy, quality, vibration fault rules, prescriptions, the Loss Ledger, reports and proof. Chat, the agents and FMEA drafting need an AI provider.

Yes. MIRA's MCP server lets other AI assistants read asset health, prescriptions, work orders and the value ledger. Each connection has exactly the permissions of the user who issued the token, never more, and write actions go through the same approval gate.

306 tool bindings across 36 agents, measured from the generated agent catalogue. The largest are Intelligent Actions (29 tools, for governed writes), Energy Optimization (23), Work Order (15), Root Cause Analysis (13) and Changeover Overrun (13). Each tool is a permission-checked read, a deterministic calculation or a governed action.

Only through governed write lanes, and never on a question alone. MIRA does not write to PLCs, SCADA or DCS. Writes to SAP or FacilityFlow are off by default and need your allow-list and a signed acknowledgement, and each action type waits for a named approver below the confidence threshold you set.