How I Use Claude and Agentic AI to Write Blog Posts Directly from My Home Assistant Setup

What Is Agentic AI, and Why Does It Matter Here?

Most people think of AI as a question-and-answer machine: you type something, it responds. Agentic AI is different. Instead of answering once and stopping, it takes a sequence of actions — calling tools, reading data, making decisions, and producing output — all within a single session and with minimal hand-holding.

Claude (made by Anthropic) operates in this agentic mode when connected to tools via the Model Context Protocol (MCP). MCP is an open standard that lets Claude connect to external systems — your WordPress site, your Home Assistant instance, a file system — and interact with them like a skilled operator would. Rather than just talking about your smart home, Claude can actually read it, reason about what changed, and then write and publish about it.

Getting Started

• A self-hosted WordPress site
• The Novamira free plugin from novamira.ai
• Claude Desktop with MCP support enabled
• A Home Assistant instance with the MCP server integration active
• A few minutes to paste the generated config into claude_desktop_config.json

The Stack: What Is Involved

1. Home Assistant — My home automation platform, running locally on Proxmox with over 2,600 entities. The source of truth for everything: cameras, sensors, energy monitoring, NVR configuration, and more.

2. Novamira (free WordPress plugin) — Novamira exposes your WordPress site as an MCP server. It gives Claude the ability to create posts, read content, and interact with Yoast SEO — through a standard interface, with no custom REST endpoints or API key juggling required.

3. Claude Desktop with MCP configuration — The client where everything comes together. Claude Desktop reads MCP server connections from a local JSON config file. Point it at Novamira and Home Assistant, and Claude can reach both.

Setting Up Novamira and the MCP Connection

Step 1: Download and Install the Novamira Plugin

Head to novamira.ai and download the free plugin. Install it via Plugins → Add New → Upload Plugin, select the zip, and activate it.

Once active, go to Novamira → Configuration and enable AI Abilities. This turns on the core tools Claude uses: creating posts, reading content, running SEO checks, and more.

Step 2: Copy Your MCP Config from WordPress

Still in Novamira → Configuration, scroll to the Claude Desktop tab. Novamira generates a ready-made JSON snippet with your unique endpoint and token already filled in. Copy it.

Open Claude Desktop and go to Settings → Developer → Edit Config to open your claude_desktop_config.json. Paste the snippet as a new entry under mcpServers. Save and restart Claude Desktop. Full walkthrough at novamira.ai/quickstart.

Step 3: Connect Home Assistant

In Home Assistant, go to Settings → Integrations and enable the Model Context Protocol integration. This provides an SSE endpoint URL and token. Add it to claude_desktop_config.json as a second MCP server entry. Restart Claude Desktop, and both connections are live.

How Claude Actually Decides What to Write

Reading Context First

When asked to write a post about a homelab change, Claude does not start writing immediately. First it queries the source — Home Assistant, a logbook entry, or a live entity state — to understand what actually happened. For the Frigate NVR post, for example, it checked which cameras were configured, which LXC container Frigate runs on, whether the Google Coral TPU was detected, and how recordings route to the NAS. This grounding in real data is what makes the post technically accurate rather than generic.

Applying a Writing Structure

Claude applies a consistent post structure based on instructions it has been given. Every post includes a keyword-rich introduction, clear H2 and H3 headings, transition words for readability (however, therefore, as a result, in addition), and a Yoast SEO meta description under 155 characters. Novamira can read and report Yoast scores back directly.

Token-by-Token Generation: What Is Happening Under the Hood

At the model level, Claude generates text one token at a time. A token is roughly a word or word-fragment. For each token, the model evaluates the full context — original instruction, tool results, everything written so far — and produces a probability distribution over what should come next. It does not think in sentences; it predicts forward, one step at a time, shaped by its training and the specific context of your request.

In practice: the more specific and grounded your prompt, the better the output. A vague prompt produces generic text. A prompt grounded in real data — your actual config and setup decisions — produces something worth publishing. The MCP connection is what makes that data available without manual copy-pasting.

Tool Use as Reasoning

When Claude uses a tool — querying Home Assistant, fetching existing post titles from Novamira — that is not just data retrieval. It is part of the reasoning process. Claude decides when to call a tool, what to ask, and how to incorporate the result. This is what makes it agentic rather than scripted: the sequence of steps emerges from the task rather than being predefined.

What a Typical Session Looks Like

A post about a recent homelab change goes roughly like this:

1. Describe what was done to Claude (e.g. migrated from BlueIris to Frigate, Coral TPU working, cameras connected)
2. Claude queries HA logbook entries and relevant entities for supporting detail
3. Claude drafts the post in chat first for review
4. Once approved, Claude calls Novamira to create a WordPress draft with Yoast meta description and focus keyphrase
5. Review in WordPress
6. Ask Claude to generate a prompt for the Featured Image. This prompt is pasted as-is in ChatGPT (free plan) to generate the image. Once generated, it can be downloaded and uploaded in the Wordpress post.
7. Publish the Wordpress Post.

The whole process takes a few minutes. The value is not speed — it is that the output is already structured, SEO-considered, and grounded in real technical detail from the actual setup.

Why This Works Well for Homelab Blogs

Most homelab posts are written after the fact, from memory, with details getting fuzzy. By connecting Claude directly to Home Assistant — the living record of what the setup looks like today — you get posts that are accurate, repeatable, and consistent. If you document changes in the HA logbook as you go, Claude can turn those notes into a publishable post almost immediately.

It also keeps the blog active. The friction of sitting down to write is often what kills posting frequency. When writing is a 10-minute agentic session rather than a two-hour blank-page exercise, you publish more. As a result, the technical record of your homelab grows alongside the setup itself.

 

The Novamira quickstart page and Anthropic MCP documentation cover each side in detail. Once the connection is live, the rest is just a conversation.