Talk to your photo library: Immich, PixelUnion and the Model Context Protocol

4 min read
Talk to your photo library: Immich, PixelUnion and the Model Context Protocol

Most AI assistants know a lot about the world and nothing about your photos. Ask one to find the pictures of your daughter’s first bike ride and it can only apologize. A small open-source project changes that for Immich, and because PixelUnion runs on Immich, it works with your PixelUnion library too.

The project is ImmichMCP, built by GitHub user barryw. It is not made by us or by the Immich team. We tested it, we like where it is going, and we want to explain what it does before you hand it the keys to your photos.

What MCP is, and why it matters

MCP stands for Model Context Protocol. Anthropic introduced it in November 2024 as an open standard for connecting AI assistants to outside tools and data. A year later it moved to the Linux Foundation, so no single company owns it anymore. Claude, ChatGPT, Gemini, VS Code, Cursor and LM Studio all support it in some form.

The idea is simple. An MCP server sits between an AI assistant and a service. It tells the assistant which actions exist, for example “search photos” or “create album”, and it carries out those actions through the service’s normal API. The assistant decides which tool to call. The server does the work.

Before MCP, every AI vendor built its own plugins for every service, in its own format. That is how lock-in starts. With one open protocol, a single server works with any assistant that speaks MCP. You can switch from Claude to a local model next month and keep the same connection to your photos. For a company like ours, that is the part that matters. We do not want your photo library tied to whichever American AI company is ahead this quarter.

ImmichMCP exposes 49 tools: smart search (“sunset at the beach”), search by date, place, camera or person, albums, tags, people, shared links, comments and uploads. In practice you can type “make an album called Iceland 2026 with all photos from February” and the assistant works out the steps. Anything destructive, like deleting photos, requires an explicit confirmation flag. The assistant cannot delete by accident without the request saying so.

How to set it up with PixelUnion

You run ImmichMCP on your own computer. It connects to your PixelUnion instance with an API key. The short version:

  1. Create an API key in PixelUnion under Account Settings → API Keys. Give it only the permissions you need. Start with read permissions.
  2. Install Docker Desktop.
  3. Add ImmichMCP to your AI client’s MCP configuration, with your PixelUnion address (https://yourname.pixelunion.eu) and the API key.
  4. Restart the client and ask it something about your photos.

The full step-by-step guide, including the exact configuration for Claude Desktop, is in our help center: Connect an AI assistant with MCP.

Now the tradeoff, because there is one. ImmichMCP runs locally, but the AI model you connect it to probably does not. If you use Claude or ChatGPT, everything the tools return goes to that provider: file names, dates, GPS locations, the names you gave to faces, and with some settings even preview images. That is data PixelUnion keeps in Europe. Once you send it to an American AI service, it is covered by their terms, not ours. If that bothers you, use a local model through a client like LM Studio. It is slower and less capable, but nothing leaves your machine.

Two more points. The API key sits in a plain-text configuration file on your computer, so treat that file like a password. And ImmichMCP is a young community project. Read the code or at least the README before you run it. Open source gives you that option. It does not do the reading for you.

What comes next

We are looking at offering an MCP endpoint as part of PixelUnion itself. That would mean no Docker, no API key in a config file, and a login you can revoke from your account settings. We have not decided yet, and we want to get the privacy side right before we build anything. If you try ImmichMCP and have opinions, tell us. That feedback shapes what we do next.

Start free with PixelUnion →

Cover photo by Vitaly Gariev on Unsplash.