Is Tollkit Extract MCP server safe?

Findings (1)

• lowUnusually long tool description (over 2,000 characters)poison.long-description

  tool get_product_details: …Find out what a product is and what it costs, from its page URL. USE THIS WHEN you have the URL of ONE specific product — a store listing, a marketplace item, a manufacturer's page — and you need its fields rather than its prose. Typical jobs: comparing the same item across several stores, tracking a price over time, checking whether something is in stock, ingesting a catalogue, or getting a product photo URL. DO NOT USE THIS FOR search-results or category pages (it returns one product, not a list), pages behind a login or paywall, or pages that are not about a product. It also cannot reach private or internal addresses. RETURNS name, description, brand, sku, price (a number), currency (ISO 4217), availability (in_stock | out_of_stock | preorder | unknown), images (up to 5 absolute URLs copied from the page, never invented) and variants. See the output schema for the exact shape. Any field the page does not state comes back null rather than guessed. HOW: the page is rendered in a real headless browser before extraction, so JavaScript-built pages work where a plain fetch returns an empty shell. COST $0.05 USDC per call over x402 on Base. No account, no API key, nothing to install, and no gas — the facilitator sponsors it. Call once WITHOUT the payment argument to get back the payment requirements, sign an EIP-3009 transferWithAuthorization with your own wallet for the quoted amount, base64-encode the x402 payload, and call again with that string as `payment`. EVERY RESPONSE SAYS WHERE EACH FIELD CAME FROM. `provenance.declared` lists the fields whose values match what the merchant published in schema.org JSON-LD — their own number, the one they publish for Google. `provenance.inferred` lists the ones a model read off the page, which might be a strikethrough price or a neighbouring product. Act on declared values; verify inferred ones if the decision matters. A URL that redirects to what looks like a DIFFERENT page — a discontinued item bouncing to its category listing — is refused free of charge rather than answered with the wrong product. Retry with `allow_redirect: true` if you want whatever the URL resolves to. A failed extraction is not charged. Repeat calls for the same URL within 300 seconds are re-served from a recent render and marked `cached: true` with the render's original `fetchedAt`. Call `try_it_free` first if you want to see real output before spending anything.…