Skip to content

Edwin MCP

A price is a claim about a moment. Most shopping tools blur that. They return a number and call it the price. Edwin returns the number with the import timestamp of the feed it came from, and its instructions to the calling model are unusually direct: never call these prices current, live or checked.

That honesty is the whole product. The rest is a competent affiliate-feed catalog across five markets, keyless, with six tools and no account.

Server type:   remote Streamable HTTP (hosted, stateless)
Endpoint:      https://thebutler.network/api/mcp
Server:        edwin 0.2.0, protocol 2025-06-18
Auth:          none, no account, no OAuth
Tools:         6
Markets:       DE, AT, CH, NL, FR, GB, US (per the tool schema)
Extension:     io.modelcontextprotocol/ui declared

The evidence tier

Keyless and driven with real offers. initialize answered 200 with no credential. tools/list returned six tools, each with a JSON schema, and the server advertises the MCP Apps UI extension.

Eleven tools/call requests were run, covering four markets and both filter bounds. Ten returned data. One returned a documented empty.

The tool surface

Tool Argument What it returns
search_products query, market, budget_max, min_price, limit Product matches with merchant count, price range and the cheapest current offer per product, each with a tracked link
get_offers pid, market Every merchant offer for one product: merchant, price, currency, stock status, source network, import timestamp
compare_products pids (two to five), market Brand, category, price range, merchant count and best offer side by side
find_alternatives product, market, optional budget and attributes Comparable products in the same market
recommend_product products or a fresh search A ranked recommendation with the reason
get_product_text product from an earlier result A short summary and key features

Every response comes in two parts: a rendered markdown table addressed to the human, and a structuredContent object addressed to the model. The markdown carries the presentation rules the vendor wants honoured, which is a tidy way to keep the offer links and the price date intact through an intermediary.

Prices, with their date attached

A German query for oled 65 zoll returned three products. The top match, an LG 65 inch OLED G6, carried a price range of 2,099 to 2,119 EUR across three merchants and a source_imported_at of 2026-11-17T22:20Z. get_offers on that product returned the three merchant rows: Coolblue at 2,099, Quelle at 2,119 and OTTO at 2,119, all in stock, all from the awin affiliate network, each with its own import time between 14:59Z and 22:20Z.

The spread is the point. Three merchants, two prices, three different import moments, and the tool does not pretend they were sampled together.

Market coverage is real. A GB laptop query returned a Dell Pro 14 at 928.09 GBP matched on exact GTIN with a match score of 1, a Samsung Galaxy Book5 at 829 GBP and a Lenovo ThinkBook at 1,047.67 GBP across two merchants. A US headphone query under 150 USD returned a Marshall Major IV at 139.99 USD sourced through the CJ network.

The refusal is the feature

A US query for running shoes under 100 USD returned zero items with empty_reason: no_match_in_market, plus a note instructing the calling model to tell the user plainly that nothing suitable was found and not to present other products as matches. A German dyson staubsauger query under 50 EUR emptied the same way.

A shopping server that returns nothing when nothing fits is worth more than one that returns a plausible near miss. This one also refuses at the argument layer: compare_products called without market returned a JSON-RPC error naming the missing argument and the expected ISO 3166-1 alpha-2 shape, instead of an empty result.

The defect worth knowing before you build on it

min_price is not honoured the way budget_max is. A German query for kaffeemaschine with min_price: 300 returned, as its first row, a De'Longhi machine at 239 EUR whose own price range tops out at 299 EUR, entirely below the floor. The same call was repeated and returned the same row. Two other rows in that response (311.82 and 338.99) did respect the floor, so the filter appears to be applied to some offers and not to the product-level match.

budget_max behaved at both ends in testing: 150 USD accepted a 139.99 USD product and 100 USD returned an empty for shoes, and a 50 EUR cap emptied a vacuum query.

The practical fix on the caller side is to read price_range.min from the response rather than trusting the filter. A floor is a request, not a guarantee.

A second quirk: a US headphone search returned the same soundcore Space 2 product twice, with two different pid values and two different tracked links, at the same 129.99 USD. Two listings, one product. Deduplicate on brand plus model before you count merchants.

Why this earns a page

For an operator, the useful part is not "an AI can shop". It is that this is a structured, timestamped, multi-market price feed an agent can query, with the affiliate disclosure built into the data rather than around it. Competitive price checks, list-price sanity checks on a supplier quote, and category research across five European markets and the US all become one call instead of five browser sessions.

The affiliate model is disclosed plainly: the offer links are tracked, and the source network (awin, cj) is a field on every offer. Anyone building on it should assume the merchant set is the merchant set those networks carry, which is why some queries empty out.

Installation

{
  "mcpServers": {
    "edwin": {
      "url": "https://thebutler.network/api/mcp"
    }
  }
}

No key, no account, no OAuth. The transport is stateless, so no session id needs to survive between calls.

Configuration

Two rules make the difference between useful results and empty ones:

  • Write the product-type word in the language of the market's merchants. German for DE, AT and CH, Dutch for NL, French for FR, English for GB and US. The schema says this and the tests bear it out: fernseher works in DE where tv finds little, and kaffeemaschine finds German machines while an English category word does not.
  • Keep brand and model names exactly as the user wrote them. Matching runs on identifiers first, so a translated brand name costs you the exact-GTIN tier.

Integration with CorpusIQ

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

Limitations

  • Prices are feed snapshots with an import timestamp. They are not a live quote and the vendor's own instructions forbid describing them as current.
  • The merchant list is the affiliate networks' catalogue, so a category a network does not carry returns an empty rather than a partial list.
  • min_price is unreliable; verify against the returned price range.
  • Duplicate product rows exist for at least one product family, so merchant counts can double count.

FAQ

Does Edwin MCP require an API key or account?

No. It answered initialize, tools/list and eleven tools/call requests with no credential and no account, and states plainly that it needs none.

Are the prices live?

No, and the server says so itself. Every offer carries a source_imported_at timestamp and the response instructs the calling model never to describe the price as current, live or checked. Quote the date with the price.

Why does a search return nothing?

Two reasons appear in the data: the query's product-type word is not in the market's merchant language, or nothing in that market fits the request. The empty comes back with a reason field and an instruction not to substitute other products.

Which markets are covered?

Germany, Austria, Switzerland, the Netherlands, France, Great Britain and the US are named in the tool schema. Driving four of them returned offers in each.

See Also

AI answers on your data. See where every number comes from.

Connect your tools and ask in ChatGPT, Claude, or Perplexity. Enterprise-grade answers in minutes. From $29.95/month.

Start free, no credit card
Contact Terms Privacy Pricing