allenporter/Home Assistant Model Context Protocol.md

## Home Assistant Model Context Protocol.md

      
    Raw
  

              Home Assistant Model Context Protocol.md
            
          
    Home Assistant Model Context Protocol integration

TL;DR

Completing these steps will let you have an LLM Powered Web scraper in Home
Assistant through the Model Context Protocol with an example of how you could make a template entity for extracting new headlines for a display.
Pre-requisites

This assumes you already know about the following:

Home Assistant
Voice Assistant with Conversation Agent (OpenAI, Google Gemini, Anthropic, etc)
Python virtual environment for running MCP server

Overview

This guide will get you an LLM powered template enttiy using MCP to fetch external web pages:

  
      Loading

  
      graph LR
    A["Home Assistant LLM Conversation Agent"] <--> |sse| B["mcp-proxy"]
    B <--> |stdio| C["mcp-fetch MCP Server"]
    C <--> |http/s| D["web pages"]

    style A fill:#ffe6f9,stroke:#333,color:black,stroke-width:2px
    style B fill:#e6e6ff,stroke:#333,color:black,stroke-width:2px
    style C fill:#e6ffe6,stroke:#333,color:black,stroke-width:2px
    style D fill:#e6ffe6,stroke:#333,color:black,stroke-width:2px

    
Install MCP Proxy & Fetch MCP Server

Install dependencies mcp-proxy and MCP Fetch Server.
$ uv pip install mcp-proxy mcp-server-fetch
Most MCP servers are stdio based (e.g. spawned by Claude Desktop) so we need to start a server to expose them to Home Assistant. We use mcp-proxy which runs an SSE server, then spawns the stdio MCP server that can fetch web pages. The proxy
server spawns the command without any environment variables, so we set our path.
$ mcp-proxy --sse-port 42783 --env PATH "${PATH}" -- uv run mcp-server-fetch
...
INFO:     Uvicorn running on http://127.0.0.1:42783 (Press CTRL+C to quit)
The SSE server is now exposed at http://127.0.0.1:42783/sse. You can set flags
to change the IP and port the proxy listens on.
Configure Model Context Protocol Integration

Important
NOTE: This integration is currently available in 2025.2.0b beta release


Manually add the integration
Set the SSE Server URL to your MCP proxy server SSE endpoint e.g. http://127.0.0.1:42783/sse. Make sure the URL ends with /sse.
The integration will create a new LLM API called mcp-fetch that is available to conversation agents. It does not add any other entities or devices.
Configure Conversation Agent


Navigate to your existing conversation agent integration and reconfigure it


Set the LLM Control to mcp-fetch


Update the prompt to be something simple for now such as: You are an agent for Home Asisstant, with access to tools through an external server. This external server enables LLMs to fetch web page contents.


Try it Out

Open the conversation agent and ask it to fetch a web page:


Prompt Engineering

Lets now experiment to use the tool to make a sensor. We should ask the model to respond more succintly, and call it from the developer tools. Here is a yaml example:
action: conversation.process
data:
  agent_id: conversation.google_generative_ai
  text: >-
    Please visit bbc.com and summarize the first headline. Please respond with
    succinct output as the output will be used as headline for an eInk display
    with limited space. 

We could even improve this by giving some few-shot example headlines, however, the model follows instructions OK already and produces a headline:
response:
  speech:
    plain:
      speech: Trump threatens Greenland and Panama Canal.
      extra_data: null
  card: {}
  language: en
  response_type: action_done
  data:
    targets: []
    success: []
    failed: []
conversation_id: 01JH2B56RMR2DABQQGAAA9X95D

Template Entity

Below is an example template entity definition based on the Model Context Protocol tool calling to fetch a web page:
template:
- trigger:
    - platform: time_pattern
      hours: "/1"
    # Used for testing
    - platform: homeassistant
      event: start
  action:
    - action: conversation.process
      data:
        agent_id: conversation.google_generative_ai
        text: >-
          Please visit bbc.com and summarize the single most important headline.
          Please respond with succinct output as the output will be used as
          headline for an eInk display with limited space, so answers must be
          less than 200 characters.
        response_variable: headline
  sensor:
    - name: Headline
      attributes:
        title: "{{ headline.response.speech.plain.speech }}"
      unique_id: "d3641cdf-aa9f-4169-acae-7f7ba989c492"
  unique_id: "272f0508-3e27-4179-9aca-06d8333874e7"

Now go out and make yourself an ESPHome E-ink display if you don't have one already!