HVAC AI Agents home

Private beta · Claude / MCP

Expose your building as an MCP server — to Claude, ChatGPT, or any MCP client.

AI-agent-ready HVAC, BMS, and home-automation. 20+ vendor adapters. Tool-call schema is open. Runs as a Cloudflare Worker, container, or local binary.

Join the waitlist →

minutes

to first agent-controlled setpoint

from npm install to a Claude tool call

20+

vendor adapters out of the box

tool-call schema is open

on-edge

on-net deployment

Worker, container, or local binary

Three modes — MCP leads here

Primary for this page

MCP server, your equipment

We adapt Modbus, BACnet, KNX, Matter, and vendor clouds into a single tool-call schema your agent can use.

Chat copilot included

Optional UI: a chat that talks to your equipment via the same MCP server. Useful for techs without an LLM client.

Autonomous on demand

Same MCP layer can be driven by a scheduler or autonomous policy if you don't want a human in the loop.

Connect Claude to a building in one step.

Point Claude at your MCP server URL. Every covered system is now a tool Claude can call.

{
  "mcpServers": {
    "hvac-ai-agents": {
      "url": "https://your-edge-bridge.hvacapi.com/mcp",
      "transport": "sse"
    }
  }
}

Model Context Protocol for HVAC

Claude MCP server — give any LLM live, write-capable access to your building's HVAC.

HVAC AI Agents ships a production-grade MCP server that exposes your BMS, VRF fleet, or thermostat network to Claude, ChatGPT, or any MCP-aware agent. Every read is sub-second. Every write is audit-logged and gated behind configurable safety interlocks.

The Model Context Protocol (MCP) is the emerging standard for connecting LLMs to external systems with structured, tool-callable interfaces. Instead of writing custom function-call glue for each AI provider, you expose your HVAC data once via an MCP server and every MCP-aware client — Claude Desktop, the Anthropic agent SDK, Cursor, a custom GPT-4o pipeline — picks it up without rework. Our server handles the protocol translation so your team writes HVAC logic, not boilerplate.

Read calls return live BMS points — zone temperatures, damper positions, chilled water valve status, fault codes — in a normalized schema regardless of whether the underlying equipment speaks BACnet, Modbus, or a proprietary brand cloud. Response time for a read call is under one second at the 95th percentile. The LLM sees a coherent, brand-agnostic building model rather than a patchwork of vendor APIs.

Write-back is the hard part, and it's where most LLM-BMS integrations fail. An LLM that can hallucinate a setpoint change that trips a safety interlock is not a deployable system. Our MCP server gates every write through a configurable interlock layer: point whitelists, value range constraints, rate-of-change limits, and time-of-day windows. Every LLM-initiated write is logged with the originating prompt, the LLM's response, the proposed value, the interlock check result, and the actual BMS write. If a write is blocked or rolled back, the audit log shows exactly why.

Authentication and tenant scoping are first-class. Each MCP client credential is bound to one or more site scopes — a credential issued for Site A cannot read or write Site B points. Multi-tenant SaaS builders can issue per-customer credentials with per-customer point whitelists. Self-hosted deployments can run the MCP server as a Docker container on the customer's LAN, with brand cloud traffic staying inside the perimeter.

Who uses the MCP server

Four teams that connected their HVAC to an LLM with our MCP server.

From solo developers to enterprise AI integrators — what they built, and what changed.

Developer building a custom Claude agent

Standing up a Claude agent that can query and adjust BMS setpoints required custom function-call schemas per brand, no common protocol, and no way to audit what the LLM actually changed in the building.

→ Connected the MCP server in an afternoon. The agent calls normalized HVAC tools — read_zone_temp, set_setpoint, list_faults — without brand-specific code. Audit log covers every tool call end-to-end.

Hours to first MCP-driven setpoint

Facilities team using Claude Desktop

Facilities staff wanted to ask Claude questions about live building conditions — 'which zones are overcooling right now?' — without waiting for an IT-managed dashboard or filing a work order with the BMS vendor.

→ MCP server connected to Claude Desktop. Staff type natural-language queries, Claude reads live BMS data and surfaces overcooling zones, fault codes, and setpoint drift in plain English. No coding required.

Zero-code LLM access to live BMS

AI integrator building multi-tenant SaaS

Multi-tenant HVAC SaaS needed to give each customer a Claude-powered HVAC assistant scoped strictly to their portfolio — no cross-tenant data leakage, no shared credential surface.

→ Per-customer MCP credentials bound to per-customer site scopes. Each tenant's Claude instance sees only their buildings. Integrator manages credential lifecycle via the MCP server's admin API.

Tenant-isolated LLM access at scale

Security-conscious on-prem deployment

Customer's security policy prohibits BMS telemetry leaving the LAN. Existing cloud-based HVAC AI vendors were non-starters. LLM-driven HVAC control was a board-level requirement with no viable path forward.

→ MCP server deployed as a Docker container on the customer's LAN. Brand cloud traffic stays local. The LLM calls the self-hosted MCP endpoint over the internal network. No telemetry transits the public internet.

Air-gapped LLM-driven HVAC control

FAQ

Claude MCP server — common questions.

  • What is an MCP server, and what can Claude do with one connected to my HVAC?

    MCP (Model Context Protocol) is an open standard that lets an LLM call external tools in a structured, auditable way. With our HVAC MCP server connected, Claude can read live zone temperatures, fault codes, damper and valve positions, and energy totals — and, with write-back enabled, adjust setpoints and reset alarms. Instead of a static report, Claude gives answers grounded in what the building is doing right now. To try it: request a sandbox credential from the form below.

  • How does the safety gate work? What prevents the LLM from making a dangerous setpoint change?

    Every write call passes through a server-side interlock layer before it touches the BMS. You configure per-point rules: allowed value ranges, max rate of change, time-of-day windows, and a mandatory human-confirmation flag for high-risk points. If the LLM proposes a value outside those rules, the write is blocked and the reason is logged. The LLM is never told the interlock configuration — it cannot reason around it. All blocked and completed writes appear in the audit trail with the originating prompt and the interlock check result.

  • What is the latency? How quickly does Claude see a BMS state change?

    Read calls return in under one second at the p95 for BACnet/IP and brand-cloud endpoints with local edge gateways. Write calls complete in one to three seconds — the additional time covers the interlock check, the BMS roundtrip, and the audit write. Polling-based brand clouds (some residential thermostat APIs) may add two to five seconds on the read path. Latency numbers for your specific integrations are available in the pilot readout. Start a pilot from the form below.

  • How do I authenticate a Claude Desktop client and scope it to one site?

    You create a credential in the MCP server admin console and bind it to one or more site scopes. The credential is a standard OAuth 2.0 client credential pair — client ID and secret — which you paste into the Claude Desktop MCP configuration. Claude Desktop presents those credentials on every tool call. The server rejects calls that reference points outside the credential's site scope. You can revoke or narrow a credential at any time without restarting the server.

  • Can I self-host the MCP server, or do I have to use your hosted version?

    Both options are supported. The hosted version (remote MCP server) requires no infrastructure from you — you get a subdomain endpoint and a credential. The self-hosted version is a Docker container that you run on your LAN or private cloud. Self-hosted deployments keep all BMS telemetry on your network; the only outbound traffic is to the LLM provider (Anthropic, OpenAI) unless you also self-host the model. Contact us for the self-hosted deployment guide and license terms.

  • Does it work with ChatGPT or other LLMs, or only with Claude?

    Any MCP-aware client works. The server implements the MCP specification, which is supported by Claude Desktop, the Anthropic agent SDK, Cursor, and an increasing number of OpenAI-based pipelines via community MCP adapters. ChatGPT's native plugin system does not yet support MCP, but OpenAI's function-calling API can consume MCP tool schemas via a thin adapter layer — we ship one. The server is LLM-agnostic at the protocol level.

  • How is this different from just calling your REST API from a function-call handler?

    Three differences matter in production. First, the MCP server handles session state, context window management, and tool-call batching — you don't rebuild that for each LLM. Second, write-back safety interlocks and audit logging are enforced at the server boundary, not in your application code, so they can't be skipped by a code change. Third, per-tenant credential scoping is a first-class server feature, not something you have to implement in middleware. If you already have function-call handlers working, migration to MCP is additive — your existing REST API still works.

  • How is the MCP server billed?

    Billing is per connected site per month, with a tiered rate that falls as site count grows. Read-only access is included in the base tier. Write-back requires the Pro tier, which adds audit log retention and interlock configuration. Self-hosted deployments are licensed annually per server instance with no per-site metering. A free sandbox credential with two read-only demo sites is available — no credit card required. Request access from the form below.

Speaks to your existing kit

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.

See full coverage →

How it stays out of your way

Secure

Sealed data plane. Per-site auth. Audit log on every setpoint touch.

Runs on the edge

Deploys at the building edge — your data doesn't leave the site to be useful.

BYO LLM

Works with Claude, ChatGPT, and any MCP-compatible client. You pick the brain.

Private beta

Get the MCP server early.

Designed for facility software companies, building tech leads, and MCP power users.