A dedicated AI workstation running Claude Desktop on Debian Linux, serving as a homelab AI companion and assistant.
Claudebox is a GMKTec K11 mini PC running Debian 13 (trixie) with Claude Desktop installed. It provides persistent, always-on access to Claude with a suite of MCP (Model Context Protocol) servers that give it real awareness of and control over the homelab environment.
The result is an AI assistant that actually knows your infrastructure — it can query live metrics, manage dashboards, read and write files, and remember context across sessions.
Claudebox uses three layers of persistent memory:
Memory MCP - Knowledge graph for structured facts: infrastructure IPs, server inventory, conventions. Shared across all chats.
Basic Memory - Markdown-based knowledge base for Claudebox-specific context accumulated through conversations. Obsidian-compatible, stored locally.
GitHub (claude-prime-directive) - The foundational reference repo. Contains communication preferences, cognitive style, infrastructure documentation, workflows, and project instructions. Loaded on demand via Desktop Commander. Acts as stable, version-controlled long-term memory that doesn't drift over time.
Together these give Claude persistent awareness of the homelab environment, working preferences, and accumulated knowledge across sessions.
RDP - Primary access method at home. Connect directly to the Claudebox desktop via any RDP client.
Guacamole - Browser-based remote access for use outside the home network. Hosted on the homelab, accessible from anywhere without a VPN or RDP client.
| Server | Purpose |
|---|---|
| Desktop Commander | File operations, terminal process execution, system commands on the Claudebox host |
| Netdata (Claudebox) | Live system metrics and alerts for the local Claudebox machine |
| Netdata (Unraid) | Live metrics and alerts for the Unraid server |
| Netdata (Atlas) | Live metrics and alerts for the TrueNAS Scale server |
| InfluxDB | Query and write time-series data in InfluxDB v2.7 on TrueNAS. Telegraf ships metrics from all hosts into per-host buckets. |
| Grafana | Dashboard management, alert rules, and querying across Netdata, InfluxDB, and Loki datasources. Also covers OnCall schedules and incident management. |
| Unraid | Array status, disk health, parity, Docker containers, shares, notifications via GraphQL API |
| TrueNAS | Pool/dataset management, SMB/NFS shares, snapshots via REST API (see note below) |
| Bluesky | Post, search, notifications, and social graph management |
| GitHub (Personal) | Repo management, issues, PRs for TadMSTR account |
| GitHub (Work) | Repo management, issues, PRs for work account |
| Memory MCP | Persistent knowledge graph shared across all chats (infrastructure IPs, conventions, facts) |
| Basic Memory | Persistent markdown knowledge base for Claudebox-specific context, Obsidian-compatible |
| Playwright | Browser automation via Firefox |
| Microsoft Learn | Search and fetch Microsoft/Azure documentation |
The TrueNAS MCP (vespo92/TrueNasCoreMCP) required a patch to http_client.py in the uv cache to fix SSL handling with the self-signed cert. The upstream bug: a custom ssl_context was passed to httpx.AsyncClient but a separate AsyncHTTPTransport was also created, which took precedence and ignored the ssl_context. Fix was to remove the transport block entirely.
Caveat: The patch is applied to the uv package cache. If uvx updates the package, the patch will be lost and will need to be reapplied.
- Device: GMKTec K11 mini PC
- CPU: AMD Ryzen 9 8945HS (8 cores / 16 threads)
- RAM: 28GB (32GB total, shared between CPU/GPU)
- OS: Debian 13 (trixie)
- Homelab monitoring and management
- Grafana dashboard creation and maintenance
- Infrastructure documentation
- Bash and PowerShell scripting assistance
- General AI assistant tasks scoped to the homelab context