homelab-asset-db-mcp

Query a Markdown-based homelab asset database via Model Context Protocol.

A tiny FastMCP stdio server that turns a plain-Markdown inventory of your homelab (servers, LXC containers, Raspberry Pis, smart-home gadgets, …) into structured MCP tools. An LLM connected via MCP can then answer "where does pve01 run", "which IP has the printer" or "what's on the smart-home VLAN" directly — no reading and grepping the Markdown file on every query.

Read-only, no hosting, no auth, no API credits. The data lives in a file you own; this repo only carries the parser + a fictional sample.

Why

Most homelab inventories live as a Markdown table in someone's notes. That's great for humans and terrible for an agent, which has to fetch the whole file and pattern-match every time. This server parses the tables once per call and exposes four precise lookups instead.

Sibling to iobroker-mcp, paperless-bulk-mcp and nanobanana-render-mcp — same FastMCP scaffolding, different domain.

Tools

All tools return JSON. Fields with no value are omitted from the response.

Setup

git clone https://github.com/McCavity/homelab-asset-db-mcp.git
cd homelab-asset-db-mcp
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Out of the box the server reads the bundled asset-db.sample.md (fictional data), so it runs immediately.

Configuration

Point the server at your own database with the ASSET_DB_PATH environment variable. Your real file stays outside this repo (it's gitignored):

export ASSET_DB_PATH=~/notes/homelab/asset-db.md

~ is expanded. If ASSET_DB_PATH is unset, the bundled sample is used. A .env file next to server.py is also read (see .env.example); values already in the environment win, so an MCP host's env config takes precedence.

Register with Claude Code

claude mcp add --transport stdio --scope user homelab-asset-db \
  --env ASSET_DB_PATH=/absolute/path/to/your/asset-db.md \
  -- /absolute/path/to/homelab-asset-db-mcp/venv/bin/python \
     /absolute/path/to/homelab-asset-db-mcp/server.py

Then /mcp lists the server and the four tools become callable in any session.

Data format

The database is Markdown. Each ## heading starts a section; the first table row after it is the header, the rest are devices. Section headings are mapped to categories in CATEGORY_MAP in server.py — headings not listed there (e.g. a "Naming Scheme" table) are treated as documentation and skipped. Columns may differ per section; each device keeps whatever columns its table defines. Both English and German section headings are recognised. See asset-db.sample.md for the full shape.

The file is re-parsed on every tool call — no caching, no restart when it changes.

Testing

python -m unittest discover -s tests -v

Covers the bundled English sample and a German fixture (so a change that breaks real-world German-headed databases fails the suite).

Known limits

• No resources (@mcp.resource) — devices without a hostname would have no stable URI. Could be added later with a slug generator.

• No caching — the file is re-parsed per call. Fine for a few hundred rows; cache + mtime-invalidate for larger inventories.

• Read-only by design — no write tools.

License

MIT © 2026 Henning Halfpap