Project information

• Category: MLOps
• Client: INSTAR Deutschland GmbH (Waletech) Shenzhen, China
• Project date: 15 Jul, 2026
• Project URL: INSTAR Cloud

Our cameras were "smart" but closed. You could only reach them through raw MQTT topics or a home-automation flow — and none of our customers' AI agents could use them, because a large language model don't speak MQTT, and handing one 700+ raw topic strings is a recipe for hallucinated commands.

So I built a FastMCP server directly on top of the MQTT control API I'd already put into the camera. It exposes six clean, domain-level tools — set a setting, read observed state, wait for a change, check camera health, pull the latest image, and list capabilities. The agent never touches MQTT syntax. The server maps a capability name to the correct topic behind the scenes, so the model reasons about "enable the red alarm area," not about `cameras/224/alarm/areas/red/enable = 1`.

The Model Context Protocol is an open standard - a single server makes the camera available to any MCP-compatible harness — Anthropic's Claude, OpenAI's Codex, and open-source agent harnesses like Hermes. I ship the integration once; every AI platform a customer already uses can drive the same hardware.

And it stays honest, which is what makes it safe. The runtime tracks state confidence — observed, known-but-stale, or unknown — so the agent reports what it actually confirmed, not what it hoped for. Credentials live in environment variables, never in the prompt or the logs, and it passes a full test suite before it ships.