AI agents for facility operations, with exactly two verbs.
Conduit is an MCP gateway that lets AI agents work on your facility through two capabilities only: Observe and Propose. Agents read the Facility Graph and draft changes; people approve them in the change process you already run.
| Stage | 03 · Operate |
|---|---|
| Reads | Facility Graph |
| Upstream | Aidni Thermal |
| Hands off to | Aidni Ledger |
| Primary output | MCP server with Observe and Propose tools |
Why this system exists
- 01
Operators want AI help with alarm triage, capacity moves and maintenance planning, but a facility is not a sandbox. An agent with broad write access to power and cooling is a risk no operator should accept.
- 02
Exposing every operation an engineer can perform as its own agent tool creates a large, hard-to-audit surface — and most of those operations are compositions of a few primitives.
- 03
Facility data is split across EPMS, BMS, DCIM and the GPU scheduler, so agents see fragments and guess at the rest.
The mechanism, step by step.
Observe
Agents query the Facility Graph and live telemetry: what exists, how it is connected, what it is doing now, and what changed. Read-only, always.
Propose
Agents draft a change — move workloads, adjust a setpoint, schedule maintenance — with its blast radius, risk, rollback and owner computed from the graph.
Approve
Every proposal lands in the customer's existing change process (ticketing, MOP, shift lead). Nothing executes without a named human approver.
Execute through your controls
Approved changes run through the systems that already own them: the scheduler, the BMS, the work-order system. Conduit never bypasses them.
Observe, propose, approve.
The agent never touches a setpoint or a job directly. It reads the graph, drafts a change with risk and rollback, and waits. A person approves, and the system that already owns the action carries it out.
| Tool | Can | Cannot |
|---|---|---|
| observe() | Read graph, telemetry, history, alarms | Change anything |
| propose() | Draft a change with risk, rollback, owner | Execute without approval |
Press run to replay an illustrative session.
- MCP server with Observe and Propose tools
- Proposal queue with risk and rollback
- Full audit log of agent reads and drafts
- Policy engine for scopes and approvers
- Model Context Protocol (MCP) clients
- ServiceNow / Jira change workflows
- GPU schedulers (Slurm, Kubernetes)
- BMS and EPMS (read)
A governed agent interface for critical infrastructure. Instead of exposing every operation an engineer can perform, Conduit gives agents two verbs, scoped by policy and approved by people, executing only through controls the operator already owns. Safety is a property of the interface, not a hope about the model.
Why this matters for the field →Bring us one site. We will show you its real capacity.
A six-week Site Assessment maps the power envelope, the cooling design and the compute plan for one facility, delivered by the engineers who build the platform.