Zalando MCP server: let your AI agent search fashion products
Wield is an MCP server that gives your AI agent three Zalando tools: a product search, a simpler HTML fallback for the same search, and an order-status tool. The searches run from your own browser with your own Zalando session. They are not working reliably yet. The search recipes passed 0 of 9 runs in our benchmark, and order status is untested.
What your agent gets
Wield exposes Zalando through wield-mcp. Your agent finds a recipe with find_recipe and runs it with run.
| Recipe | What it does | Login | Effect | Status (7 Nov 2026) |
|---|---|---|---|---|
zalando.de/search-products | Searches fashion products with optional size and price filters | No | Read | OK at 12:05, in-page and offline check. Benchmark: 0 of 9 runs |
zalando.de/search-products-html | Fallback: reads the first 24 product cards of the catalogue page | No | Read | OK at 12:05, in-page and offline check. Benchmark: 0 of 9 runs |
zalando.de/order-status | Shows the status of your recent Zalando orders | Yes | Read | Untested. Needs a logged-in recording |
The main search returns the product name, brand, price, original price, sale flag, sizes in stock, badges and product link. The HTML fallback returns fewer fields: name, brand, price, reference price, discount and link.
Use the fallback when Zalando rotates the query hashes the main search depends on. Then the main recipe fails with expect_failed.
How a request flows
- You ask your agent: "Find sneakers in size 42 under 80 euros."
- The agent calls
find_recipe, thenrunwithzalando.de/search-products. - wield-mcp hands the run to the Wield extension through a local bridge.
- The extension sends two requests from a zalando.de tab. They carry your own browser session.
- The results return to your agent. Nothing passes through wield.run.
What leaves your computer
Each search sends two requests to www.zalando.de, from your own browser tab, one search at a time. Wield logs the method, host, path, status and timing on your computer, and never logs bodies or query values. Telemetry is off by default. Passwords and cookies never leave your computer.
Zalando terms, in plain words
Zalando's terms forbid automated extraction of content. The search recipes make two requests per search, one search at a time, and you start each one. Zalando's bot protection (Akamai) scores automated traffic. In our runs, an automated test browser was tagged as aggressive-crawler traffic, so keep searches human-paced. Buying stays on zalando.de. Read Zalando's terms yourself.
What does not work yet
- Benchmark misses. Both search recipes passed 0 of 9 runs. The cloud browser agent passed 0 of 9 as well, and bot protection blocked it in all 9 runs. We do not present this as a working result.
- Live check and benchmark disagree. The live check on 7 November 2026 passed, but none of the 9 benchmark runs did. Our test machine cannot verify the search automatically, because Akamai returns soft failures to command-line requests.
- Order status is untested. The recipe needs a logged-in recording before it can be used. Its field names are not verified yet.
- Fallback is smaller. The HTML fallback reads the first 24 products and returns fewer fields than the main search.
FAQ
Is there an official Zalando MCP server?
Not that we know of as of 7 November 2026. Wield is not a Zalando product. It runs public recipe files for Zalando's search.
Which AI apps work with it?
Any MCP client. npx -y wield-mcp init sets up Claude Desktop, Claude Code, Cursor and Codex. Other MCP clients work through the standard stdio config. See AI apps that work with Wield.
Does it need a login?
Product search does not. Order status does, and that recipe is untested, so it is not ready to use.
How fast is it compared with a browser agent?
In our benchmark, as of 7 November 2026 (preliminary, still running), the Wield search recipe passed 0 of 9 runs. A Browser Use agent driving Claude Sonnet 5.5 also passed 0 of 9 runs, and bot protection blocked it in all 9. Speed is not a fair comparison for this site yet. The full method and raw logs are on the benchmark page.