Swervie

Swervie MCP server

One scored answer per product, across every retailer, for AI agents.

Remote MCP server · Streamable HTTP · no auth · no API key · free

Developer docs · Tool reference · Connect a client · Response shapes · Data handling

---

https://us-central1-swervie-prod.cloudfunctions.net/mcp

Swervie is a cross-retailer product catalog built for assistants. Ask a normal shopping question and get one consolidated, scored answer per product: live prices, comparable specifications, review-derived ratings with verbatim evidence quotes, and a buy link for every seller. Instead of eight duplicate listings and a stale listicle.

• Consolidated. Listings of the same product across retailers collapse into one row. White-label clones sold under different brands are grouped too.
• Structured. Every category has its own spec schema; specs are extracted from manufacturer pages, retailer listings and spec tables, with provenance.
• Scored. Each product carries a quality score and a value score, plus review-derived sub-ratings backed by quotes from real reviews.
• Deterministic and compact. No LLM in the serving path. Answers are precomputed pipeline output: ~4k tokens, sub-second for a full shelf.
• Interactive in hosts that support MCP Apps. Results render as a shelf with sort controls, seller-by-seller offers and the review evidence (io.modelcontextprotocol/ui extension); plain JSON everywhere else.

Coverage today: 300+ categories, from robot vacuums and televisions to espresso machines, mattresses and hiking boots, across hundreds of sellers. The live list is at swervie.xyz/llms.txt.

Quick start

Claude Code

claude mcp add --transport http swervie https://us-central1-swervie-prod.cloudfunctions.net/mcp

Cursor, Windsurf, VS Code, Claude Desktop, ChatGPT, Gemini CLI and others: see docs/clients.md. Every config is the same one line, the URL above, with no headers and no secrets.

curl

MCP=https://us-central1-swervie-prod.cloudfunctions.net/mcp

H=(-H 'Content-Type: application/json' -H 'Accept: application/json, text/event-stream')

Handshake: server info, capabilities, agent instructions

curl -s -X POST $MCP "${H[@]}" -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"curl","version":"1.0"}}}'

A whole shopping answer in one call

curl -s -X POST $MCP "${H[@]}" -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"search_products","arguments":{"query":"65 inch OLED TV under $1500","limit":3}}}'

More in examples/: a Node client using the official SDK, a Python client, and the curl script above.

Tools

Every tool is read-only and annotated (title, readOnlyHint: true, openWorldHint: true), so hosts can run them without a confirmation prompt. The server provides data only: it writes nothing on the user's behalf, and neither its instructions nor its results direct the model. Full input schemas and descriptions, generated from the live server, are in docs/tools.md.

A typical conversation

1. "Find me a 65-inch OLED TV under $1,500" → search_products → ranked shelf with prices, buy links and the commission disclosure.
2. "Which of those is better for gaming?" → get_products → refresh rate, VRR, HDMI 2.1, and what reviewers say about input lag.
3. "Is the Sony A95L any cheaper anywhere?" → lookup_product → every seller offer, plus its rank on the shelf and higher-scoring alternatives.

The server ships instructions on initialize that describe each tool and the fields its results carry, so a client needs no prompt engineering.

User signature (optional, no PII)

Every tool accepts an optional user_signature: an opaque, stable, pseudonymous id that gives each shopper their own rate-limit quota (instead of sharing one per session or egress IP) and attributes their buy-link clicks. The recommended recipe:

user_signature = "sv1-" + sha256hex("swervie/v1|" + id).slice(0, 24)

where id is the user's most stable identifier, normalized (lowercase, trimmed; email preferred, with +tags stripped). Never send the raw identifier. If none is available, use "sv1r-" + a UUID kept for that user, or omit the field. A value that changes every session counts as a new shopper each session. The server hashes any raw email or phone number that arrives by mistake before anything is stored.

Limits, affiliate links, and privacy

• Rate limits. Anonymous tier: 60 requests/min and 2,000/day per client (signature, session, or IP), with a higher per-IP ceiling sized for shared agent-platform egress. Partner keys with their own quota are available for platforms: contact info@plum-app.com.
• Buy links are affiliate links. Every response that carries buy_url also carries a disclosure string worded for the shopper. Commissions never affect rankings, scores, or which products are shown.
• No account, no checkout, no payments. Purchasing hands off to the retailer's own page.
• What is stored. Tool name, query, category and product ids per call, the signature when present, and rate-limit counters. Nothing the user owns, and nothing written on the user's behalf.

Details in docs/data-handling.md. Policies: privacy · terms · disclosure.

What is in this repository

This repository is the public home of the Swervie MCP server: documentation, the registry manifest (server.json), client configuration, examples, and brand assets for directories. The serving code runs as a Google Cloud Function inside Swervie's main monorepo and is not published here. The hosted endpoint is the supported way to use it. Issues and feature requests are welcome.

Server version is reported by initialize (serverInfo.version). Protocol version 2025-06-18. Transport: Streamable HTTP, stateless.

Changes: v1.6.0 (2026-10) removed the two non-destructive write tools (record_selection, submit_feedback) and rewrote the server instructions and result notes as descriptions of the data; the server is now read-only end to end.

License

Documentation, examples and assets in this repository are MIT licensed (© Plum Applications, Inc.). Use of the hosted service is governed by the terms.