Private beta · Datacenters
AI agents that watch every CRAC, AHU, and chilled-water loop. Designed for 0.05-0.15 PUE improvement. Full audit trail of every cooling decision.
Join the waitlist →0.05-0.15
designed PUE improvement
across mixed legacy + modern rooms
7+ days
early warning on CRAC failure
predictive baseline learning
sidecar
to your existing DCIM
Modbus / SNMP / BACnet native
Three modes — autonomous leads here
Primary for this page
Modulates supply temp + airflow + chilled-water against IT load + weather + tariff.
On-call asks: 'why did CRAC-3 alarm at 02:14?' Copilot answers + flags the next likely failure.
Expose room telemetry to Claude or your in-house model via MCP.
Data center HVAC
HVAC AI Agents normalize telemetry across mixed-brand CRAC fleets — Liebert/Vertiv, Stulz, Schneider, Mitsubishi, Daikin — so colo operators and edge-DC teams get real-time PUE rollups, hot/cold-aisle drift alerts, and capacity headroom forecasts without replacing the cooling infrastructure they already own.
Measuring true PUE in a multi-brand CRAC environment is harder than it looks. Each manufacturer surfaces power draw, supply air temperature, and chilled water valve position through different protocols — some via BACnet, some via SNMP, some through proprietary cloud APIs. The result is a patchwork: facilities sees one number, the DCIM sees another, and the capacity planner is working off a spreadsheet that was current six weeks ago. The agent collapses that patchwork into one normalized telemetry stream, so PUE is a live metric instead of a monthly calculation.
On top of that unified stream, the agent runs continuous monitoring. PUE and WUE roll up by zone, row, and site in real time. Hot/cold-aisle drift — the first sign that airflow is mismatched to IT load — triggers an alert within minutes, before SLA impact or thermal throttling. Capacity headroom is modeled per row and per zone against current IT load, with a 30-day forward forecast based on provisioning trends. If a row is within 10% of its cooling ceiling, the capacity planner knows before the next cage fill.
Integration is designed for the data center stack, not retrofitted from commercial BMS tooling. The agent reads BACnet/IP and BACnet/MSTP for mechanical systems, SNMP for PDU-adjacent cooling units, and standard REST APIs for DCIM platforms — Sunbird, Nlyte, FNT, and Schneider EcoStruxure IT — without requiring a forklift upgrade to any of them. Edge sites with intermittent WAN are supported via a lightweight gateway that tolerates connectivity loss and syncs on reconnect. Write-back to chilled water valves is available but strictly opt-in, with configurable safety interlocks on every point.
Where it pays off
Four patterns we see repeatedly across colo, hyperscaler regional, and edge-DC teams.
Colo ops director
Mixed-brand CRAC fleet — Liebert on one hall, Stulz on another, Schneider in the newest pod. Three dashboards, no unified PUE number. Monthly PUE is assembled by hand from BMS exports.
→ One normalized PUE dashboard with per-row breakdown, updated in real time. Aisle drift flags within minutes. Monthly PUE report generates itself.
PUE live, not monthly
Capacity planning manager
Forecasting cooling headroom against IT load growth takes weeks of manual modeling across rack manifests, CRAC datasheets, and DCIM snapshots. Cage-fill decisions lag by a month.
→ Per-row capacity headroom modeled against current IT load and provisioning trends, with a 30-day forward forecast. Capacity planners see which rows have headroom before committing cage fills.
30-day headroom forecast
Hyperscaler regional ops lead
BMS fires hundreds of alarms per shift. No ranking by severity. Aisle drift — the leading indicator of a hot-spot SLA breach — is buried in noise.
→ Agent ranks alerts by severity and surfaces aisle drift as a first-class signal. Regional team acts on thermal anomalies minutes after onset, before servers throttle or SLA triggers fire.
Minutes to aisle-drift alert
Edge-DC site lead
Intermittent WAN at edge sites means most monitoring tools go dark during outages. When connectivity returns, the gap in telemetry makes post-incident analysis unreliable.
→ Lightweight edge gateway buffers telemetry locally during WAN loss and syncs on reconnect. Monitoring and alerting stay live at the site level regardless of connectivity. Post-incident log is complete.
Zero telemetry gaps on reconnect
Your agent's outdoor signal
Wet bulb and enthalpy are the levers your HVAC actually responds to. Pulled live from open-meteo for your location.
data: open-meteo.com · location: Ashburn, VA
Live data unavailable
FAQ
The agent reads IT load from PDU or BMS points and cooling power draw from each CRAC unit, then computes PUE continuously — not monthly. Because it normalizes telemetry across brands, a Liebert CW valve and a Stulz EC fan contribute to the same rollup. Improvement comes from two levers: flagging aisle drift before it forces CRACs to overcool, and trimming chilled water supply temperature on mild days (opt-in, with interlocks). Most deployments see 0.05–0.15 PUE improvement at the 90-day mark. Start a pilot from the form below.
Both. The agent reads mechanical cooling systems (CRACs, AHUs, chilled water plants) over BACnet/IP and BACnet/MSTP. SNMP is used for PDU-adjacent cooling units and any network-managed infrastructure that exposes cooling metrics via MIBs. If a device supports neither — some older Stulz units use proprietary serial — we have integration connectors for the most common proprietary protocols. Check /coverage for the full protocol list, or contact us with your specific model numbers.
Yes. The agent integrates with Sunbird DCIM, Nlyte, FNT Command, and Schneider EcoStruxure IT over their standard REST APIs. Integration is read-only by default: the agent pulls rack power and space data for capacity modeling without writing to your DCIM. If your DCIM already has partial telemetry, we reconcile agent data with what DCIM has rather than duplicating it. No DCIM upgrade is required. Contact us with your DCIM version and we will confirm the integration path before the pilot.
The agent has production integrations for Liebert/Vertiv (Liebert DS, CW, and XD series), Stulz (CyberAir and WIB), Schneider (InRow and Uniflair), Mitsubishi (eCO2 and R2 series), and Daikin (Applied Magnitude and Alerion). For each brand, we normalize key points — supply air temperature, return air temperature, compressor power draw, and chilled water valve position — into a common schema. Coverage details and model-level support status are on /coverage.
Read-only by default. Write-back is opt-in per site and per point. When enabled, the agent may adjust chilled water supply temperature setpoints and CRAC fan VFD speeds within configurable bounds, always subject to your safety interlocks. Every write is logged with the reason, the before/after value, and a rollback path. Points you don't approve for write-back stay read-only permanently — the agent never escalates its own permissions. Most clients start read-only for 30–60 days, then selectively enable write-back after validating alert quality.
Edge sites use a lightweight gateway appliance deployed on-site. The gateway collects telemetry locally, buffers it to local storage during WAN outages, and syncs to the cloud platform on reconnect. Alerting logic runs on the gateway, so aisle-drift and overtemp alerts fire locally even when the WAN is down — delivered via on-site SMS relay or local email server if configured. Hardware footprint is a single 1U or mini-PC class device. Most edge sites are running within a day of gateway installation. Ask us about our edge-DC deployment kit.
DCIM cooling modules model cooling capacity from the design spec — rated tons, design airflow — and report IT load from IPMI or PDU data. They rarely ingest real-time CRAC telemetry from the units themselves. The HVAC AI Agent reads actual operating telemetry from every CRAC: live valve position, actual fan speed, measured supply and return air temperatures. The result is capacity headroom based on what the CRACs are actually doing today, not what they were designed to do. The two tools complement each other; the agent does not replace your DCIM.
Per site (data hall or facility), with tiering by site count and average installed cooling tonnage. No per-CRAC, per-rack, or per-point metering. Pricing includes DCIM integration, edge gateway hardware for qualifying sites, and capacity headroom reports. A single-site pilot is a fixed-fee engagement. Multi-site rollouts move to an annual contract. Request a quote from the form below — pricing depends on your CRAC count and DCIM platform.
Still have a question?
Type a question. We'll route it to the right page or capture it for our team.
Carrier, Trane, Daikin, Mitsubishi, LG, Lennox, York, Samsung — 20+ HVAC, home-automation, and BMS brands.
63 brands across 3 categories — HVAC (31), Home Automation (18), BMS (14). Protocols: BACnet, KNX, MQTT, Matter, Modbus, REST, WebSocket, Z-Wave, Zigbee.
How it stays out of your way
Sealed data plane. Per-site auth. Audit log on every setpoint touch.
Deploys at the building edge — your data doesn't leave the site to be useful.
Works with Claude, ChatGPT, and any MCP-compatible client. You pick the brain.
Private beta
Designed for colo, edge IT, and enterprise datacenter ops.
Required for the site to work. Bot protection (__cf_bm cookie set by Cloudflare) and
form rate-limiting (one-way HMAC of your IP, 1-hour expiry). Cannot be disabled.
Anonymized traffic stats via Cloudflare Web Analytics (cookieless). Helps us understand what pages people visit.
Save-and-resume for long forms (localStorage). If off, you'll need to complete forms in one sitting.