--- name: gyazo description: Search, read and summarise a person's Gyazo captures from the command line with the `gyazo` CLI - screenshots and phone photos, their OCR text, the application or page they came from, and where they were taken. Use when asked what someone saw, captured, read, bought, visited or worked on, when a question needs evidence from their screen history, or when a Gyazo image ID, a gyazo.com URL or a collection needs to be read. --- # Gyazo CLI `gyazo` reads a Gyazo account: screenshots, phone photos, their OCR text, the application and page a capture came from, and the coordinates and address of a photo. It caches everything it reads under `~/.cache/gyazocli`, so repeated questions about the same period are answered locally. Check `gyazo config get me` first when a token might be missing; it prints the account or exits non-zero. Set one with `gyazo config set token `. Every command exits non-zero on failure, so `&&` chains and `set -e` behave. ## Reading captures ```bash gyazo ls --limit 10 # most recent, newest first gyazo ls --date 2026-08-30 # a local day; also yyyy-mm and yyyy gyazo ls --today gyazo ls --hour 2026-08-30-14 # one hour, from the cache only gyazo ls --photos # shorthand for has:location gyazo get # one capture in detail gyazo get --ocr # just the OCR text gyazo # same as get gyazo # same as get gyazo ./screenshot.png # an existing file uploads instead ``` Add `-j`/`--json` to any of these when the output is going to be parsed rather than read. Add `--no-cache` when the answer must come from the API. ## Searching `gyazo search ` takes Gyazo's own query language. The operators below were confirmed against the live API; see [references/search-syntax.md](references/search-syntax.md) for the measurements behind them and the ones that do not exist. ```bash gyazo search "お好み焼" # OCR text, title and description gyazo search "has:exif" # photographs, not screenshots gyazo search "address:広島" # where a photo was taken gyazo search "date:2026-08-30" # a day, a month or a year gyazo search 'app:"Gyazo Android"' # the application it came from gyazo search "ocr:Wi-Fi has:exif" # terms are ANDed gyazo search "address:広島 OR address:京都" # capital OR gyazo search "has:location -app:Chrome" # leading - negates gyazo search "has:exif" --page 2 --limit 50 # 20 per page by default ``` Three things worth knowing before composing a query: - **An operator Gyazo does not know returns nothing**, rather than falling back to a text search. A guessed operator looks exactly like "no such captures". - **`has:exif` is how to narrow to photographs.** The application does not tell them apart: `app:"Gyazo Android"` includes screenshots and screen recordings from the same phone. - **There is no coordinate or radius search.** Search by place with `address:`, which matches the address in any language and matches postal codes too. ## Summaries and rankings ```bash gyazo summary # the week to yesterday, day by day gyazo summary --date 2026-08-30 gyazo stats --days 30 # one markdown report gyazo apps --date 2026-08 # what applications, most used first gyazo domains --today gyazo tags --date 2026 gyazo locations --date 2026-08-30 ``` The ranking commands take `--date`, `--today`, `--limit`, `--max-pages`, `--json` and `--no-cache`. Only `stats` takes `--days`. ## Collections ```bash gyazo collection gyazo collection https://gyazo.com/collections/ gyazo collection --sort captured # added | created | captured gyazo collection --anonymous # read a public one without the token ``` A collection ID and an image ID are both 32 hex characters and cannot be told apart on their own, which is why `gyazo ` reads it as an image. Pass the `/collections/` URL when the ID is a collection. A collection larger than 100 images is reported as truncated: the public endpoint returns the first 100 and cannot page. ## Filling the cache ```bash gyazo sync --days 7 # yesterday back through 7 days gyazo sync --date 2026-08 # a whole month gyazo sync --query "has:exif OR has:location" --max-pages 20 ``` `sync` covers yesterday backwards and never today, because today is still happening. For anything from today use `ls --today`, `search`, or a ranking command with `--today`. `--query` fills the cache from a search instead of the listing, which is how to gather one kind of capture without walking past everything else: photographs are a small fraction of a day's screenshots. Put any date range inside the query (`date:2026-08`, `since:... until:...`) rather than in `--date`, which `--query` refuses. Budget about 40 seconds per page of 100 captures that are not cached yet. ## Answering questions with captures - A day is a local day. `--date 2026-08-30` means that date in this machine's timezone, which is also how the cache is laid out. - Prefer `search` with an operator over `ls` plus filtering. The API does the work, and `ls --date` over a wide range walks many pages. - OCR text is noisy: it comes from screenshots at whatever resolution, and `locale` is often `und`. Treat it as a hint, not a transcript. - `get --objects` prints detected objects, but the API no longer returns the field it reads, so it exits non-zero with "Object annotations not found" on every capture tested. Use the OCR text instead. - **Do not turn a capture into a claim it does not support.** A product page or a cart is interest; an order confirmation or a payment receipt is a purchase. Say which capture the conclusion rests on. - A phone photo carries coordinates and a reverse-geocoded address in `metadata.exif_normalized` and `metadata.exif_address` of the `--json` output; the top-level `exif_normalized` is always null. Screenshots carry neither. ## Serving the same data over MCP `gyazo --mcp-server` runs the same functionality as a Model Context Protocol server over stdio, for a client that speaks MCP rather than shell. Everything above is the shell path.