AI agents for AV operations
MCP servers that give AI assistants live AV fleet data, used by Teams bots and Claude Desktop, with read-only tools by default and a confirmation gate on every action.
Problem
The fleet data already existed in the AV Device Manager, but getting at it meant opening the app or a vendor portal and knowing where to look.
The AV team and wider staff wanted to ask simple questions in plain English, such as “is this room working?” or “which devices at this client are offline?”, from the tools they already use.
Constraints
- Actions on real customer hardware must be deliberate. An assistant must never reboot a device or change its configuration because it misread a request.
- The same tools had to work from more than one interface: Microsoft Teams for staff, and Claude Desktop for engineers.
- No vendor credentials could be exposed to the model or to the client application.
Architecture
The tools live in MCP servers that sit in front of the existing fleet store and vendor APIs. The AI clients see tool definitions, never credentials.
What I built
MCP servers for the AV fleet. Written in JavaScript, each server exposes tools with clear names and typed inputs that give assistants live device and client data:
- Read-only tools such as device status, room health and firmware versions. These run immediately.
- Action tools that change something on a device. These sit behind a confirmation gate: the tool first returns a summary of exactly what it will do, and the action runs only after a person explicitly confirms it.
One-click Claude Desktop extensions. The servers are packaged as Claude Desktop extensions (.mcpb bundles), so an engineer can install them in one step.
Teams bot agents. Agents built in Copilot Studio and Azure AI Foundry use the same MCP servers, so staff can check room status in plain English from inside Teams.
Training. I ran training sessions and live demos of these tools for colleagues and management, backed by written guides the team now uses.
Results
- Staff can check room status in plain English from Teams, and engineers from Claude Desktop, using the same tools.
- No action runs on customer hardware without an explicit human confirmation.
- The team is trained on the tools and works from written guides.
What I’d do next
- Scope tools by role, so service desk staff get read-only tools and engineers also get action tools.
- Test tool-call accuracy against a fixed set of real questions before each change ships.
- Surface every confirmed action in the existing audit log view, with who confirmed it and when.