Hi everyone,
I built a Home Assistant addon that exposes a local directory to Claude
(or any LLM agent) via Model Context Protocol (MCP).
Why?
Claude was already connected to my HA via HA-MCP and to my Keenetic
router via my own Keenetic-MCP addon. It could see devices, read logs,
manage automations.
But between sessions — it forgot everything. Every chat started from
scratch.
Then Andrej Karpathy published his LLM wiki idea: a personal knowledge
base in plain markdown files, maintained by the LLM itself. No vector
databases, no RAG. Just a folder of .md files the model reads and writes
directly. 16M views on X — clearly struck a nerve.
I decided to implement this for my smart home.
What it does
The addon runs an MCP server inside Home Assistant that exposes a
directory (e.g. an external USB drive at /media/VAULT) to Claude via
Model Context Protocol.
Combined with HA-MCP and Keenetic-MCP, Claude can now:
- Read current state of all devices and automations via HA-MCP
- Read network info via Keenetic-MCP
- Write structured wiki pages about your setup to the vault
- Read them at the start of the next chat — and already know the context
The addon auto-creates the vault structure and a CLAUDE.md instruction
file on first run.
Setup overview
- Connect an external USB drive to your HA server
- Format it as ext4 with label VAULT
- Install Samba NAS addon (dianlight) for auto-mounting
- Install this addon and configure your token
- Expose port 3100 via your router
- Add the MCP endpoint to claude.ai as a custom connector
Status
Tested and working on amd64
Beta branch available for aarch64/armv7 (Raspberry Pi) —
needs testers!
Links
GitHub: GitHub - st412m/ha-filesystem-mcp · GitHub
Karpathy’s LLM wiki: llm-wiki · GitHub
Would love feedback, especially from anyone who tests the beta on
Raspberry Pi or other ARM hardware!