| name | search-by-image |
| description | Reverse-image search on AliExpress — upload a product photo to AliExpress's own visual search (camera icon) via the browser and return IL-context product cards. For items that are hard to name (industrial parts, fittings, odd shapes) where keyword search fails. |
Search by Image (reverse-image search)
Some products can't be found by words — you have a photo (or a competitor's listing
image) but no reliable name, and keyword search returns the wrong category entirely
(e.g. searching "shelf bin / parts organizer" surfaces jewelry organizers, not the
industrial louvre pick-bin you actually want). AliExpress's own visual search is
excellent at these. This skill drives it: upload the image → read the IL-context
result cards.
When to use
- You have an image of the product (photo, screenshot, or a saved listing image)
and keyword search isn't surfacing it.
- The item is a hard-to-name commodity (industrial bins, brackets, fittings, an odd
shape) — the complement to
search-by-synonyms (which attacks the naming problem
from the words side; this attacks it from the picture side).
Why the browser, not an API
The official AliExpress Open Platform API (affiliate / dropshipping) does not cleanly
expose consumer reverse-image search. AliExpress's site visual search is the
reliable route, and this plugin already drives the site in the user's own browser, so
it's the natural home.
- Primary (this skill): visual search on
he.aliexpress.com, driven in the
user's own Chrome (mcp__claude-in-chrome__*) — see
$CLAUDE_PLUGIN_ROOT/reference/browser.md for the route order. Gateway Playwright
is the unattended fallback.
- Fallback (no browser / batch): third-party image-search APIs —
Apify
freecamp008/search-by-image-aliexpress, omkarcloud/aliexpress-scraper —
take an image URL and return matches as JSON. They cost/limit per call and are
outside the IL locale, so treat their prices as indicative and re-check landed cost
via fetch-listing.
Inputs
image (required) — local file path to the product photo, or an image URL.filters (optional) — same as search-aliexpress (choice default true,
freeshipping, rating4plus, ship_from); applied on the results page after the
visual search returns.defining_spec (optional but recommended) — the one attribute that separates the
target from look-alikes (usually a dimension). Visual search returns
visually-similar items across the whole size range, so filter on this.max_results (optional, default 20).
Method (Chrome first)
Do the locale handshake first ($CLAUDE_PLUGIN_ROOT/reference/browser.md) so
results come back in ₪. Load mcp__claude-in-chrome__file_upload in the same
ToolSearch call as the rest of the browser tools — you will need it at step 2.
- Open the entry point. Navigate to
https://he.aliexpress.com/. The visual
search lives behind the camera icon inside the main search bar. Anchor on role
/ aria-label (icon class names rotate — never hard-code them); the control is an
image-search / camera button adjacent to the search input. Clicking it reveals a
drag-drop zone with a hidden <input type="file">.
- Upload the image. Hand the local
image path to that file input —
mcp__claude-in-chrome__file_upload in Chrome, browser_file_upload under
gateway Playwright. If given a URL instead of a local path, download it to a temp
file first, then upload. On success the page navigates to the image-search results
(URL contains image-search / an uploaded-image token).
- Wait + read cards. Wait for results to render, then read product cards with the
same defensive query as
search-aliexpress
(a[href*="/item/"] → url / title / price / badges), de-duped by item id.
- Apply filters + spec. Apply any requested chips (Choice default). Drop cards
that fail
defining_spec — this is essential for visual search, which happily
returns the 30×20 version of a 60×40 product.
- Hand off. Shortlist →
fetch-listing (landed cost) → hunt-pricing-anomaly
(per-SKU / ₪-per-litre) on any multi-size listing.
Selector discovery (do once, record it)
The camera/image-search control and the file input are the only image-specific
elements, and they are not yet in reference/selectors.md. Validate them live
the first time, preferring aria-label / role anchors, then write what worked into
that file with a date. If the in-search-bar camera isn't present for the account,
the image-search results page can be reached directly and the file input targeted
there.
Output format
Image: <path or url> Defining spec: <e.g. No.3 pick bin ≈ 24×14.5×13 cm>
Route: claude-in-chrome
Visual-search results URL: <url>
Locale verified: c_tp=ILS
Result count: <N> (passing defining spec: <K>)
1. <title> ₪<price> badges:[Choice] <url>
2. …
Dropped-on-spec: <count>
Next: `ship-options-il` on the top picks, then `fetch-listing` for landed cost.
Composition
- Pairs with
search-by-synonyms: image search finds the product family; synonyms +
the spec filter refine and widen it. Run whichever the input gives you (photo vs.
words), then cross-check.
- Feeds
fetch-listing and hunt-pricing-anomaly.
Validation checklist
- The uploaded image produced an AliExpress image-search results page (URL /
heading confirms visual search, not a keyword query).
- Locale/currency verified (prices in ₪).
defining_spec applied and the dropped count reported (visual search without a
spec filter is almost always too broad).- At least the top picks were handed to
fetch-listing for landed cost.
