Best MCP Servers for Home Lab and Self-Hosting

Home Assistant ships an MCP server as a normal integration. Add "Model Context Protocol Server" in Settings, and your instance serves MCP at /api/mcp, with OAuth or a long-lived token. The agent can only see and control entities you have exposed to Assist, and it cannot change your Home Assistant configuration. That makes it the safest server on this list.

For Claude Code, Home Assistant's docs give an add-json command with OAuth. Claude Desktop needs a public URL or mcp-proxy for a LAN-only instance.

The power-user option: 87 tools covering automations, scripts, dashboards and configuration, far beyond the official integration. Very active (release 8.5.0 on September 16, 2026, about 4.9k stars) but unofficial. It has a read-only HTTP mode. Use that until you trust your setup, because the full mode can rewrite automations. If you are building voice control, also see our guide to Google Home with MCP and AI agents.

There is no official Proxmox MCP server. This fork is the best maintained (release 0.5.22 on September 26, 2026) and adds what matters for safety: a read-only policy and approval for high-risk actions like stopping or deleting a VM. It covers VMs, LXC containers, backups and snapshots. Run it with uvx proxmox-mcp-plus or Docker.

Create a Proxmox API token with a limited role (PVEAuditor for read-only) rather than using root. The original canvrno/ProxmoxMCP has not been updated since February 2025. Our Proxmox VE guide covers setting up roles and tokens.

Portainer's own server (release 2.45.1, September 2, 2026) with a read-only mode. It goes through Portainer's API and permissions instead of the raw Docker socket, which is the safer design. One catch: the server's minor version must match your Portainer version.

Two uses here. First, it runs other MCP servers each in their own container, with managed secrets, which is a tidy way to host all the servers on this list on your home server. Second, its catalogue includes Docker management servers. For direct container control, the community mcp-server-docker (GPL-3.0) manages containers and compose stacks.

Be clear about the risk: access to the Docker socket is effectively root on that machine.

For k3s and Talos home clusters. Supports read_only = true, selectable toolsets, Helm and KubeVirt. An alternative is mcp-server-kubernetes (MIT, with a non-destructive mode): claude mcp add kubernetes -- npx mcp-server-kubernetes. Use a kubeconfig for a limited service account.

The most useful monitoring server. The agent can search dashboards, run Prometheus and Loki queries through Grafana, and read alerts, which is exactly what you need for "why was the NAS slow last night?" Official and very active (1.6.0 on September 25, 2026). Run with uvx mcp-grafana, a Grafana service account token with Viewer role, and --disable-write.

Now hosted under the Prometheus organisation. Query metrics directly, with destructive admin tools switched off unless you opt in. Handy if you run Prometheus without Grafana, or want the agent to explore metrics you have not built dashboards for.

From iXsystems itself, which is good, but labelled a research preview and "not recommended for production" (0.0.6, July 2026). Fine for reading pool health and dataset usage on a home NAS. Do not let it touch pools or shares yet.

claude mcp add truenas -- truenas-mcp --truenas-url 192.168.1.10 --api-key YOUR_KEY

There is no official Unraid server yet. This one (formerly jmagar/unraid-mcp) uses Unraid's GraphQL API and is actively maintained, but its README says plainly that it is not read-only. Use an API key with limited permissions.

31 tools for searching the library, checking what is playing and managing items, with --read-only and --disable-destructive flags and confirmation prompts. Updated September 2026. More on the server itself in our self-hosted media servers comparison.

Covers Plex plus Sonarr, Radarr and Trakt in one server: 46 tools, 58 with writes enabled. "Add the new season of this show and tell me when it is ready" works. Keep writes off unless you want the agent adding downloads. Standalone *arr MCP servers exist but are all very small projects; treat them as unvetted.

n8n has an instance-level MCP server at /mcp-server/http. You opt workflows in one by one, and since 2.13 the agent can also build and edit workflows. This turns your existing automations into tools: "run the backup workflow" or "build a workflow that pings me when the NAS is full". Turn it off entirely with N8N_DISABLED_MODULES=mcp. New to it? Read what is n8n and how to self-host n8n.

The highest-risk category. For OPNsense, OPNSenseMCP hides write tools with OPNSENSE_READ_ONLY=true. For pfSense, pfsense-mcp-server wraps the REST API package in 332 tools with confirmation for destructive actions, which is a lot of context for most agents. We would only use either in read-only mode, for questions like "what blocked this device last night?"

Tailscale has no official MCP server, though its Aperture product can proxy other MCP servers across your tailnet. Community servers can list devices and manage ACLs; keep them read-only, since an ACL change can cut off access. Cloudflare's official server (https://mcp.cloudflare.com/mcp) handles DNS, Tunnels and Workers well if you expose services through Cloudflare Tunnel.