MCP Inject

An OpenAI API proxy server that injects tools into the context using MCP.

The use case this was made for was to add web search to Home Assistant's Voice Assistant. Normally when Model Context Protocol is used, the client is responsible for all of these:

• Telling the LLM what tools are available
• Running the tools on behalf on the LLM
• Including the results from running those tools in the context

The purpose of this tool is to do all of the above so the client doesn't have to. This is especially useful in cases where the client doesn't have MCP features (e.g. many Home Assistant LLM integrations).

Configuration

Set the MCP_INJECT_BACKEND_URL environment variable to the base URL of your model server. For example, http://127.0.0.1:8000/v1. URL paths like /chat/completions will be suffixed to the provided URL.

If running in Docker, it's recommended to set the TZ environment variable as well, as that impacts the current time tool. For example, TZ=America/New_York.

Starting the proxy server

If you have uv installed, you can start the server like so:

cd mcp-inject
uv sync
uv run fastapi run app/main.py

Without uv, you can start the server like so:

cd mcp-inject
python3 -m venv .venv
source .venv/bin/activate
pip install .
fastapi run app/main.py

Since the proxy server is a FastAPI server, you can pass FastAPI flags, like port to bind to a specific port:

fastapi run app/main.py --port 8001

Example setup

1. Set up an OpenAI compatible model server.
2. Run MCP Inject and point it at the model server.
3. Run an OpenAI API client, for example, the Local LLMs Home Assistant integration. Configure it to talk to MCP Inject.

Now that Home Assistant integration will be able to respond to prompts with the help of web search.

Currently included tools

• duckduckgo-mcp - Web search through DuckDuckGo. This search backend was chosen for simplicity as it doesn't require an API key. This MCP search also provides a tool to fetch the content of a URL.
• Getting the current date and time - This is handy, but it's really here because it was the simplest tool to integrate during initial development.