The gap between AI agents and physical infrastructure just got smaller. A developer behind the open-source homelab-monitor project has integrated a read-only Model Context Protocol (MCP) server directly into their monitoring stack. This addition allows AI agents, such as those powered by Claude, to query real-time system metrics without requiring manual data pasting or complex API integrations.

Bridging the Visibility Gap

The creator noted a specific pain point: while agents could execute commands, they lacked situational awareness of the hardware state. The new MCP server, running on port 9810, exposes the same data the dashboard displays. This includes granular details like per-card GPU memory allocation, process-level RAM usage, and energy costs derived from power integration. The setup is streamlined with a single command: claude mcp add --transport http homelab http://YOUR-HUB:9810/mcp.

Security and Scope Constraints

Crucially, this implementation is strictly read-only. The agent can observe the lab's statusβ€”identifying which model is occupying a 3090 GPU or checking disk treemaps on a Raspberry Piβ€”but it cannot restart services or modify configurations. The developer emphasized that while the dashboard retains control features, the MCP surface is designed to keep agents from becoming responsible for critical system actions. All data remains on the local LAN, ensuring no information leaves the house when paired with local models like Ollama.

Key Takeaways

  • Seamless Integration: The MCP server rides along with the existing monitoring container, requiring no extra installs on remote nodes.
  • Granular Data Access: Agents can query specific metrics like per-card GPU splits (e.g., 22.5 / 22.1 / 18.8 GB) rather than pooled totals.
  • Open Source & MIT Licensed: The project is available on GitHub, with the MCP feature enabled by default and toggleable via enable_mcp=0.
  • Local-First Privacy: The architecture ensures all telemetry and agent interactions remain within the local network.

The Bottom Line

This is a pragmatic step toward agentic infrastructure management. By standardizing read-only access via MCP, we're giving agents the context they need to reason about hardware without the risk of unintended side effects. It's the kind of utility that makes homelabs truly conversational.