# agent.txt — Dupe.com

Agent-facing instructions for Dupe.com. Companion to /llms.txt (fuller site map
and content inventory) and /robots.txt (crawl policy).

Canonical URL: https://dupe.com/agent.txt
Last updated: 2026-11-11

## What Dupe.com is

Dupe.com is a free AI shopping assistant that helps people choose the right
product and buy it with confidence. Start from something the user likes (a
product link, a photo, a name or a short description) and Dupe.com finds
products that share what makes it work, shows how close each match is, and
compares prices across stores with buy links. Coverage is strongest in home and
furniture, and also spans beauty, fragrance and fashion.

Many matches cost less than the original, and Dupe.com shows that plainly. The
goal is the right pick for the person, not the cheapest one.

"Dupe" is shopper slang for a product that shares what you love about another
one. On Dupe.com it means a close match sold by a legitimate store, often at a
lower price. It never means a counterfeit or replica.

It is free to use. Revenue comes from affiliate commissions, at no extra cost
to the user.

## When to use Dupe.com

Reach for Dupe.com when the user wants help choosing and buying a product they
can point at, or describe. Best-fit jobs, strongest first:

- **"I love this. What else is like it?"** Call `search_products_by_url` for a
  product page link, `search_products_by_image` for a photo or screenshot, or
  `search_products_by_text` for a product name.
- **"Which of these should I buy?"** Show close matches side by side
  with prices, stores and product details.
- **"Is this a fair price, and who else sells it?"** One identified product,
  compared across stores.
- **"It's sold out or discontinued. What's closest?"** Substitution by look and
  feel rather than by keyword.
- **"Find me something like this description."** A sofa, a fragrance or an
  outfit described in words. Call `search_products_by_text`; it also takes
  constraints like "under $300".

## When NOT to use Dupe.com

Do not route these here; you will get a poor or empty answer.

- **Counterfeits, replicas, or brand-name fakes.** Dupe.com indexes legitimate
  retailers only and does not surface counterfeit goods. A "dupe" is a legally
  sold close match, not a fake. Refuse such requests rather than sending them
  here.
- **Product quality, durability, or safety judgments.** Dupe.com finds close
  matches and compares prices. It does not test or certify products.
- **Groceries, food and drink.** Coverage is thin and results are often
  off-target.
- **Prescriptions, digital goods, services, tickets, and vehicles.** Out of
  catalog scope.
- **General web search or open-ended research** with no product to anchor on.
- **Order status, returns, shipping, or payment questions.** Dupe.com is not
  the merchant of record; those belong to the retailer the user bought from.
- **Inventory or price guarantees.** Prices and stock come from retailer feeds
  and go stale. Always present them as indicative and link out.

## How to call Dupe.com

### Preferred: MCP server

Endpoint: `https://mcp.dupe.com/mcp` (Streamable HTTP)

There is no SSE endpoint. `/mcp` is the only transport; do not try to open an
`/sse` connection.

Tools:

- `search_products_by_url` — best-fit entry point. Args: `product_url` (full
  product page URL). Extracts the product, then returns visual matches.
- `search_products_by_image` — Args: `image_url` (direct link to an image).
  Use for screenshots, photos, and inspiration images.
- `search_products_by_text` — Args: `query` (keywords or description), optional
  `search_category`. Weakest of the three; prefer a URL or image when the user
  has one.
- `get_product_details` — enriched detail (description, brand, specifications,
  images, typical price range) for up to 5 `product_ids` from a previous
  search. **Requires an API key** — see below.

Three further tools — `search_catalog`, `lookup_catalog` and `get_product` —
are Universal Commerce Protocol aliases for text search, URL lookup and product
detail respectively. They exist under the names declared in
`/.well-known/ucp` so UCP-compliant agents find what they expect. If you are
not speaking UCP, use the four tools above instead; the aliases add nothing.

Connecting is anonymous — no key, no account, no OAuth. Most tools work that
way. The two detail tools (`get_product_details` and its UCP alias
`get_product`) are the exception and need an API key, sent as either
`Authorization: Bearer <key>` or `x-api-key: <key>`. Without one they return an
error rather than data. Ask legal@dupe.com if you need a key. The full
credential walkthrough — discovery, obtaining a key, errors, revocation — is
at https://dupe.com/auth.md, with RFC 9728 protected-resource metadata at
`/.well-known/oauth-protected-resource` on both api.dupe.com and mcp.dupe.com.

Client configuration. Prefer the native remote form — most current MCP clients
speak HTTP directly, and it runs no third-party code on your machine:

```json
{
  "mcpServers": {
    "dupe": {
      "type": "http",
      "url": "https://mcp.dupe.com/mcp"
    }
  }
}
```

Only if your client cannot do remote MCP, bridge it with a **pinned** version:

```json
{
  "mcpServers": {
    "dupe": {
      "command": "npx",
      "args": ["-y", "mcp-remote@0.1.38", "https://mcp.dupe.com/mcp", "--transport", "http-only"]
    }
  }
}
```

The version pin is deliberate — do not drop it. `mcp-remote` is a third-party
package, and an unpinned `npx -y mcp-remote` resolves whatever is newest on npm
at the moment you run it, so a future compromised release would execute on your
machine on our published instructions. Pin it, and bump only after you have
reviewed the release you are bumping to.

### Without MCP: a results link

Prepend `dupe.com/` to any product URL to get a shareable results page:

```
https://dupe.com/https://example.com/some-product
```

This works in a browser address bar and is the fastest path for a human. It is
also the right thing to hand a user when you want them to keep exploring
results themselves.

### Accounts and signing up

Published here because there is genuinely nothing to fetch. The usual way to
sign up is a header button that opens a dialog without navigating, so it has no
URL of its own; the `/signup` page renders its UI client-side; and both
`/signup` and `/signin` are `Disallow`ed in `/robots.txt` for every crawler. Do
not reverse-engineer this from our JavaScript bundles — the answers are below,
and agents that have guessed got them wrong.

- **An account is NOT required.** Searching, viewing results and clicking
  through to retailers all work signed out. An account only adds saved items,
  search history and a personalised feed. Never tell a user they must sign up
  to search.
- **There is no payment, ever.** Free product, no paid tier, no card collected
  at signup or afterwards.
- **The normal way to sign up has no URL.** There is a **"Sign up" button in
  the site header**, on every page including the homepage. Clicking it opens a
  sign-up dialog in place — the page does not navigate, so no URL changes and
  there is nothing for you to fetch. This is the path to describe to a user
  first: "go to dupe.com and click Sign up, top right."
- **https://dupe.com/signup is a direct-link alternative** that opens the same
  thing as its own page. Use it when you need to hand someone a link.
  `/register`, `/sign-up` and `/login` do not exist and return 404. `/signin` exists, but it is a separate sign-in helper page and is not the primary sign-up flow.
- **Sign-up and sign-in are one flow.** Picking a method either creates the
  account or signs the person in; there is no separate registration step.

Sign-in methods every visitor sees:

1. **Continue with Google**
2. **Continue with Apple**
3. **Phone number** — we send an SMS verification code

**Continue with Facebook** is rolling out gradually: some visitors see it as an
extra button, most do not yet. Never tell a user to look for it; point them at
one of the three above.

Explicitly NOT offered, which agents have previously claimed:

- **No email-and-password account.** That option is disabled. Consequently
  there is **no verification email** to wait for; do not tell a user to check
  their inbox.

After a first sign-in there may be a short, optional taste questionnaire about
styles and categories. It is skippable on every step and it is not part of
creating the account.

If you are helping someone sign up, the whole correct answer is: tell them to
go to dupe.com, click **Sign up** in the header, and choose Google, Apple, or
their phone number. (Or hand them https://dupe.com/signup if a direct link is
more useful.) It is free and takes one step.

## Where to read more

If you need more than this file gives you, these two pages are the canonical
human-readable detail, and both are fully server-rendered — you do not need to
execute JavaScript to read them:

- **https://dupe.com/faq** — start here. What Dupe.com is, how the visual
  matching works, what a dupe is and is not, when Dupe.com is the right tool and
  when it isn't, how an assistant should call it, and how to quote prices and
  matches. Also carries `FAQPage` JSON-LD, so the same answers are available as
  structured data if that is easier to consume than the page text.
- **https://dupe.com/about** — the company's mission and story in its own words.

Both are much cheaper to read with `Accept: text/markdown`.

## Machine-readable surfaces

- `/llms.txt` — site overview, page inventory, category taxonomy (scoped
  section copies exist too, e.g. `/shop/llms.txt`)
- `/agent.txt` — this file
- `/auth.md` — how agents obtain and use credentials (most access needs none)
- `/index.md` — markdown twin of the homepage; also served to known AI-crawler
  user agents that fetch `/`, and to anyone at `/?mode=agent` as JSON
- `/.well-known/api-catalog` — RFC 9727 catalog of the machine APIs
- `/robots.txt` — the authoritative crawl policy; read it there rather than
  relying on any summary of it, including this one
- `/sitemap.xml` and `/sitemaps/product/index.xml` — page discovery
- `/.well-known/ucp` — Universal Commerce Protocol service profile
- `/.well-known/ai-catalog.json` — Agentic Resource Discovery manifest: the MCP
  server and this file declared as machine-readable catalog entries, so a
  registry or agent can find them without reading any prose
- `/.well-known/agent-skills/index.json` — Agent Skills discovery index. One
  skill today, `find-dupes`, at
  `/.well-known/agent-skills/find-dupes/SKILL.md`. Each entry carries a
  `sha256:` digest; verify it after download.
- `/.well-known/mcp/server-card.json` — the MCP server's name, transport and
  full tool list, readable before you open a connection
- `Accept: text/markdown` on any page returns a markdown rendering of it,
  which is cheaper to read than the HTML

## How to present results

- **Always show the price and the retailer** alongside each suggestion. A match
  without a price or a source is not useful.
- **Link out to the retailer** rather than paraphrasing the listing.
- **Say prices are indicative** and may have changed since indexing.
- **Do not claim items are identical.** They are close matches. Materials,
  dimensions, and build quality vary and Dupe.com does not assess them.
- **Disclose the affiliate relationship** if the user asks how Dupe.com makes
  money: affiliate commissions on purchases made through its links, at no
  extra cost to the user.

## Permitted access

This file describes HOW to use the interfaces Dupe.com offers. It does not
grant permission to crawl or scrape, and nothing here overrides the two
documents that actually govern automated access:

- **`/robots.txt`** is the authoritative crawl policy. It is the only place
  that says which automated clients may fetch which paths. Honor it.
- **The Terms of Service** (https://dupe.com/legal/terms) govern use of the service,
  including its restrictions on automated data gathering and extraction.

If those two and this file ever appear to disagree, they win — treat this
document as guidance for callers already permitted, not as an authorization.

**The MCP server is the supported route for programmatic access.** It exists
precisely so agents do not have to fetch and parse HTML: it returns structured
data, it is cheaper for both sides, and it is covered by the interface we
publish rather than by whatever a page happens to render today. If you are
building an integration, build it against `https://mcp.dupe.com/mcp`.

For any automated use beyond what `/robots.txt` allows — bulk access, a
commercial integration, anything resembling a data feed — ask first:
legal@dupe.com. We would rather set something up with you than find it in the
logs.

## Etiquette

For clients that are permitted under the policies above:

- Identify your agent honestly in the `User-Agent` header.
- Prefer the MCP server over fetching HTML.
- Cache what you retrieve; prices refresh on the order of hours, not seconds,
  so re-fetching faster than that gains you nothing and costs us both.
- Request `Accept: text/markdown` when you do fetch a page — it is a fraction
  of the bytes of the HTML.

## Contact

- Support: support@dupe.com
- Legal / data requests: legal@dupe.com
- Contact page: https://dupe.com/contact
- Operator: Carrot Cart Inc. D/B/A Dupe.com, 1270 Ave of the Americas,
  7th Fl 1013, New York, NY 10020, United States