ESXi MCP

ESXi MCP bridges Claude to VMware's virtualisation stack, letting you query VM states, start or stop instances, and pull real-time performance metrics without leaving your chat window. It speaks to both standalone ESXi hosts and vCenter, using VMware's pyvmomi library under the hood and exposing everything through MCP's tool interface. I'd reach for this when managing a homelab or small-scale virtualisation cluster where you want conversational access to VM lifecycle tasks. Ask Claude to list all VMs on a host, check CPU and memory usage, or power-cycle a stuck instance. The SSE support means you can stream performance data in real time, which is useful for capacity planning or debugging resource contention without opening the vSphere client. For homelabbers running a handful of ESXi boxes, this turns Claude into a lightweight ops assistant. The install is manual: pip install the dependencies, then point the server at your ESXi or vCenter endpoint. You'll need to pass credentials and host details, which the repo's README covers but doesn't automate. No environment variables are pre-configured, so you're writing your own connection strings. The upside is flexibility; the downside is you're on your own for secrets management. Quirks: this is a community project, so expect rough edges. The repo doesn't ship as a PyPI package, meaning you clone and run locally. Documentation assumes you know your way around VMware's API surface. If you've never touched pyvmomi or don't already have ESXi running, the learning curve is steep. Also, the 'none' package status means updates are manual pulls from GitHub. Who shouldn't bother: anyone without an existing VMware environment, or those expecting a polished, zero-config experience. If you're running Proxmox, Hyper-V, or cloud VMs, this won't help. And if you're managing production vCenter at scale, you probably want proper monitoring tools, not an LLM wrapper. But for the homelab crowd who already speak VMware and want to experiment with agentic VM management, it's a clever proof of concept. I've used it to script VM snapshots via Claude and pull resource stats during load tests. It works, but you'll need to read the code to trust it.