Free during launch: every app and tool is unlimited — no account, no card. Purchases open soon (then $9 once). Download · Free browser tools
GUI Hub
Guide · Photo Library Search (AI descriptions)

Let Claude search your photo library by what’s in the pictures — offline

Find the photo of your passport, a receipt, a whiteboard — or the video of the kids at the beach — among thousands of local files. A small AI model describes each one on your computer; Claude searches the descriptions.

Free · no account · 3 free uses of every tool · also for Mac (Intel), Windows, Linux

Somewhere in 20,000 photos is the picture of your passport, the receipt you need for a return, or the whiteboard from last month’s meeting. Your AI can look at a photo you attach — but it can’t open your library, and uploading all of it to find one picture isn’t practical or private.

What you ask

Index my Pictures folder, then find the photo of the receipt from the hardware store.

What happens

  1. photos_index runs a small vision model (SmolVLM-500M) on your computer and writes one line per photo — what kind of image it is, what’s in it, any readable text, and the date — into an index file. A few seconds per photo; run it again any time and it continues where it stopped, skipping photos it already described.
  2. photos_videos adds your videos to the same index: nine frames from each one are described together, so “the video of the kids at the beach” can be found too.
  3. photos_search finds the best matches for what you describe, and Claude shows you the paths, dates and descriptions — or reads the index itself for trickier questions (“which photos from March show the car?”).

JPG, PNG, WebP, TIFF, BMP and GIF work everywhere; iPhone HEIC photos work on a Mac. The descriptions come from a small model: great for “receipt”, “passport”, “dog on a beach” or a sign’s text; don’t expect it to recognise specific people.

Safety: treat what it finds as untrusted

Documents and especially emails can contain text written to manipulate an AI (“ignore your instructions and…” — prompt injection). GUI Hub hands everything it reads to your AI fenced and labelled as untrusted data, and emails are marked as third-party content. That lowers the risk but can’t remove it: models can still be fooled. So don’t give the same AI session both your private files and tools that can send data out (web requests, shell commands with network access) without keeping permission prompts on. GUI Hub’s own tools only read local files and write local results — they never send anything anywhere.

Under the hood

Runs llama.cpp (MIT), the official release, with Hugging Face’s SmolVLM-500M model (Apache-2.0) — both checked by SHA-256. Unofficial wrapper — credit to both projects.

Set it up

In Claude Code: /plugin marketplace add https://guihub.cc/.claude-plugin/marketplace.json then /plugin install gui-hub@gui-hub — or install GUI Hub Essentials (version 0.22 or newer). Every tool works 3 times free; $9 once unlocks everything.

Try it now

Free download — use it 3 times free. $9 once unlocks everything. Prefer to ask your AI? GUI Hub also works as an MCP server for Claude, Cursor and others.

Other things you can do