I Made My Family Dinner AI's Problem — Part 2: Now the Whole House Is In On It

Source: Dev.to

Background

In Part 1 I built a meal‑planning system with Claude that used a preference file, reusable skills, and pushed groceries to my Kroger cart. It worked great—for me. The rest of the family, however, was still outside the loop.

We already had a family shopping list in Home Assistant, accessible via a kitchen tablet and a phone app. My daughter could add applesauce when we were low, and my husband could add bread. The meal planner and the Home Assistant list lived in completely separate worlds. Every week I had to generate a grocery list from the meal plan, manually compare it with the family list, and merge the two before ordering.

Challenges

All of our home services—Home Assistant, n8n, Nextcloud—run on a server inside our house and are not exposed to the public internet. This keeps things simple and avoids security headaches.

Claude Code runs locally, but the cloud‑based Claude Cowork cannot reach into our private network; Anthropic’s servers cannot resolve internal hostnames or route traffic to our house. I explored options such as Claude’s custom connectors and Cloudflare Tunnel, but each required exposing our home server to the internet, which I wanted to avoid.

Solution Architecture

The key insight was that the Kroger MCP server from Part 1 already runs on my Mac, which is on the home network and can reach everything inside the house. I only needed a way for Claude to communicate with n8n without exposing anything externally.

Two components made this possible:

1. n8n’s “MCP Server Trigger” node – turns any n8n workflow into an MCP endpoint.
2. mcp-remote npm package – added to the Claude Desktop configuration (the same config file from Part 1) to forward calls to the n8n endpoint over the LAN.

The data flow looks like this:

Claude Desktop (my Mac)
    │
mcp-remote (also on my Mac)
    │ HTTPS over LAN — never leaves the house
    ▼
n8n (home server)
    │
Home Assistant shopping list

“Remote” here means the other side of my Wi‑Fi, not the internet.

I built two n8n workflows behind this bridge:

• Read the Home Assistant shopping list
• Remove items after they’ve been ordered

Implementation Details

MCP Bridge Setup

1. Install mcp-remote on the same Mac that runs Claude Desktop.

   npm install -g mcp-remote

2. Add a new entry to the Claude Desktop config (the same file used in Part 1) pointing to the n8n MCP endpoint, e.g.:

   {
     "name": "home-bridge",
     "url": "https://192.168.1.42:5678/mcp"
   }

3. Configure n8n with an MCP Server Trigger node that listens for the incoming request and then calls Home Assistant’s REST API to fetch the shopping list.

Workflow Logic (simplified)

flowchart TD
    A[Claude requests shopping list] --> B[n8n MCP Trigger]
    B --> C[Call Home Assistant API]
    C --> D[Return list to Claude]
    D --> E[Claude merges with meal plan]
    E --> F[Kroger MCP adds items to cart]
    F --> G[Optional: n8n removes ordered items]

• The first workflow returns the current Home Assistant list to Claude.
• Claude merges this list with the weekly meal‑plan groceries, deduplicates, and sends the combined list to the Kroger MCP server.
• The second workflow watches for order confirmations and removes those items from the Home Assistant list, keeping everything in sync.

Outcome

What started as “I need to plan dinner” has become a system the whole family uses without knowing the AI is involved. My daughter and husband interact only with the Home Assistant shopping list on the kitchen tablet or their phones. On Thursday, Claude silently reads that list, merges it with my meal‑plan groceries, deduplicates, and pushes everything to Kroger.

The system now consists of three MCP servers working together:

1. n8n bridge – reads the family’s Home Assistant shopping list.
2. Meal‑planner skills – generates the week’s grocery needs.
3. Kroger MCP – adds everything to the cart.

The family’s “we need milk” and my “we need 4 lb of chicken breast” end up in the same Kroger cart—no manual merging required. The AI is invisible plumbing, not a product the users interact with directly.

Conclusion

Your home network is an underrated AI platform. If you self‑host anything, it can probably talk to Claude through this same bridge pattern—just one config entry pointing at your LAN.

The best AI systems are invisible to most of their users. Let the system grow from actual use, not from planning. I fixed dinner, then connected the shopping list, and the broader ecosystem emerged organically.

What’s running on your home network that could benefit from a bridge?