Offers MCP

Offers MCP

Every offer. One call.

Deep purchase intelligence for AI agents, applications, and humans. Point it at a product (free text like Sony WH-1000XM6 black, or a GTIN/UPC/EAN barcode), add an optional US zip code, and one call scans merchants in parallel to return live prices, true cost, delivery options, and trust insights. Robust, private, efficient results for agents, applications, and humans to make informed purchase decisions.

Robust. Built to be trusted:

• Live prices with true landed cost (price + shipping, the number that actually matters)
• Delivery windows for a given zip code
• Merchant trust scores (no pay-to-rank), with untrusted listings filtered
• Confidence scores on the product match and on every offer, so your agent knows when to assert and when to hedge. When we're not sure, the response says so instead of guessing
• A computed best trusted offer and fastest trusted offer, plus the full option list. The final call stays with your user
• Provenance on everything: every offer names its source and the moment it was checked (checked_at, cache status)

Private. Nobody watches users shop: no per-user tracking, no raw IPs stored; queries stay private: never sold, shared, or used for retargeting (details in Privacy).

Efficient. Shop many stores in seconds: one input fans out to multiple sources at once and comes back normalized into one structured contract you can build against. Repeat lookups return instantly. Use it for a cross-vendor price check from the store aisle or programmatic lookups inside your own application. Same tool call, same answers.

Free during v0. No signup. No API key. No commitment.

Quick start

The hosted server speaks streamable HTTP at:

https://offers-engine.dexter-works.com/mcp

Claude Code

claude mcp add --transport http offers https://offers-engine.dexter-works.com/mcp

Claude Desktop / claude.ai (no terminal needed)

1. Open Settings → Connectors (paid plans).
2. Click Add custom connector.
3. Name it Offers, paste the URL https://offers-engine.dexter-works.com/mcp, and save.
4. In a new chat, make sure Offers is enabled in the tools/connectors menu, then just ask your shopping question.

ChatGPT

1. Settings → Apps & Connectors → Advanced settings → enable Developer mode (Plus/Pro/Team plans).
2. Back in Connectors, choose Create / add a custom connector.
3. Name it Offers, MCP server URL https://offers-engine.dexter-works.com/mcp, authentication: none.
4. In a chat, enable the Offers connector from the tools menu and ask away.

Cursor

Add to .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{ "mcpServers": { "offers": { "url": "https://offers-engine.dexter-works.com/mcp" } } }

VS Code (GitHub Copilot)

Add to .vscode/mcp.json:

{ "servers": { "offers": { "type": "http", "url": "https://offers-engine.dexter-works.com/mcp" } } }

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json:

{ "mcpServers": { "offers": { "serverUrl": "https://offers-engine.dexter-works.com/mcp" } } }

Any stdio-only MCP client

Use mcp-remote as a bridge:

{
  "mcpServers": {
    "offers": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://offers-engine.dexter-works.com/mcp"]
    }
  }
}

Client UIs move fast. If a menu path above is stale, open an issue and we'll fix it.

Then ask your assistant something like:

Use Offers MCP to check if I'm getting a good deal on the Sony WH-1000XM6 at $398. I'm in 83702.

Naming the tool matters: assistants have their own web search and will often reach for it first on shopping questions. Saying "use Offers MCP" (or enabling only this connector for the chat) makes sure your question hits live structured offers instead of search snippets.

REST API

The same engine is available over plain HTTP:

curl "https://offers-engine.dexter-works.com/api/search_offers?q=Sony%20WH-1000XM6%20black&zip=83702"

Tools

Tool	Input	Output
search_offers	query (text or GTIN), zip (optional, 5-digit US)	Normalized offers + best/fastest trusted offer

More tools (identify_product for GTIN/URL/image resolution and variant families) are on the roadmap.

Response shape (abridged)

{
  "product": { "name": "Sony WH-1000XM6…", "brand": "Sony", "resolution_confidence": 0.9 },
  "offers": [
    {
      "merchant": { "name": "Best Buy", "trust": { "score": 95 } },
      "price": { "amount": 398.0, "list_price": 449.99, "discount_pct": 11.6 },
      "fulfillment": {
        "landed_cost": 398.0,
        "delivery": { "zip": "83702", "min_days": 2, "max_days": 2, "basis": "merchant_stated" }
      },
      "variant_match": { "confidence": 0.98 },
      "provenance": { "checked_at": "2026-08-11T19:42:03Z", "cache": "miss" }
    }
  ],
  "summary": {
    "best_offer_id": "dw:off_…",
    "recommendation": "Best Buy at $398.00 with free shipping is the best trusted deal."
  }
}

Sources, limits, and expectations

• Sources (quality data, not screen-scraping): we work through official merchant and catalog partner programs and structured data sources so results stay accurate rather than brittle. Today offers come from Google Shopping, with retailer-direct integrations (Best Buy, Walmart) rolling out; GTIN inputs are resolved via Icecat and UPCitemdb. Each response lists exactly what was queried in meta.sources_queried, and where sources disagree or fall short, the response says so.
• Latency: a cold query takes roughly 5–15 seconds (live upstream lookups); repeated queries hit a short-lived cache and return in milliseconds. Configure generous tool timeouts.
• Rate limits / fair use: free during v0. About 10 fresh lookups per visitor per day (cached queries are unlimited and don't count); over-cap requests get an honest capacity message, and capacity resets daily (UTC). Heavy or commercial use: contact us first. The endpoint may evolve; breaking changes will be announced in this repo.
• Agent contract: machine-readable guidance for agents (how to interpret confidence, freshness, trust, and affiliate fields) lives at offers-engine.dexter-works.com/agents.md.
• Feedback tool: report_feedback(query, verdict, note?) files a result report straight from your assistant; every report is read.
• REST and MCP return the same JSON shape (the MCP tool wraps it as text content).
• Note when curling /mcp directly: responses are SSE-framed (event: message / data: {...}) per MCP streamable HTTP; MCP clients handle this for you.

Privacy

We built Offers for ourselves first, and we treat your searches the way we'd want ours treated.

• What we see: the product you searched, the zip code you optionally provide, and coarse technical context (which kind of client connected, country/region). We don't track individuals or build per-user shopping profiles, and anonymous use requires no account at all.
• What we don't keep: raw IP addresses are never stored; unique usage is counted with a hash that changes daily and can't be reversed.
• What it's used for: one thing: making results better (match accuracy, merchant coverage, speed).
• What we will never do: sell your data, share it with advertisers or data brokers, or use queries for retargeting. Full stop.
• Merchants can't watch you compare. All price comparison happens on our servers; a merchant learns about a shopper only at click-through to their site to buy. Some outbound links may earn us an affiliate commission at no cost to you, and commissions never influence ranking (trust scores and landed cost decide, and that logic treats every merchant the same).

Questions or want something deleted? hello@dexter-works.com.

Status & support

Early v0: US consumer goods, a small set of sources done honestly. If a field can't be trusted, it's null rather than guessed.

Found a bad result, wrong price, or a bug? Open a GitHub issue. Bad-match reports with the exact query are especially valuable right now.

Built by Dexter Works.