Claude, connected to your storefront.
Seven tools over MCP, running on your own machine. No account and no API key to start.
Can agents buy from comfykoalas.com?
Yes, with one thing worth fixing. Here is how that store reads to an agent:
Exchange illustrative. The three verdicts are what the live endpoint returned for that store on 2026-08-11.
{
"mcpServers": {
"cresva": {
"command": "npx",
"args": [
"-y",
"@cresva/acp-mcp"
]
}
}
}Paste into claude_desktop_config.json and restart Claude Desktop.
Seven tools, and nothing that is not one of them
Rendered from the same list the package is tested against, so this page cannot advertise a tool the server does not register. Verified against the published 0.1.1 build on 2026-08-11.
discover_storefront
Find a brand's agent-readable storefront and what it publishes.
search_products
Search a storefront's catalogue by query.
list_products
List a storefront's products with their offer fields.
get_product
Read one product, including price, currency and availability.
negotiate_price
Ask a storefront what it will do on price, where it supports it.
get_trust_score
Read the storefront's published trust signals.
compare_products
Compare products across a storefront on the fields agents read.
7 tools in @cresva/acp-mcp 0.1.1
Ask Claude to scan any store
Every other tool takes a Cresva brand id, which means you already have to be a customer. This one takes a URL.
scan_store
Ships in 0.2.0Scan any public store URL and report whether agents can find it, understand it and buy from it. No brand id, no key.
Illustrative, not a transcript
“Can agents buy from comfykoalas.com?”
Claude calls scan_store, and answers from the verdict: the offer publishes price, currency and availability together, so the buy axis passes, while the product markup is microdata rather than JSON-LD, which is the thing worth fixing. Those two facts are what the endpoint actually returned for that page on 2026-08-11.
Ships in 0.2.0. The endpoint behind it is live and needs no key, and the tool is written, but the build on npm today is 0.1.1, which carries the seven storefront tools above and not this one. Installing now gets you those seven; scan_store arrives when 0.2.0 is published.
What you would actually ask
Three shapes the seven tools support. Claude chains them itself, which is the part that makes a tool list feel like a capability.
Look a storefront up
What does this brand publish for agents?
discover_storefront answers what an agent can actually see: whether there is an agent-readable storefront at all, and what it exposes. It is the call every other one depends on.
Read the catalogue
Find their nappies under twenty dollars, in stock.
search_products and list_products walk the catalogue, and get_product returns one item with the fields a purchase decision needs: price, currency, availability.
Compare before recommending
Compare these three on price and availability.
compare_products puts items side by side on the fields agents read, and get_trust_score reports the trust signals the storefront publishes about itself.
A process on your machine, calling one API
Nothing sits between the first two: they talk over stdio on your own machine. The third hop is a normal HTTPS request, and it is the only thing that leaves.
Locality, stated precisely
The server runs on your own machine over stdio. It is not a hosted endpoint and there is no Cresva session in the middle. It does reach the network: each tool call is an HTTPS request to api.cresva.ai, which is where the storefront data lives.
Keys
No API key is needed to start. Tools run anonymously and are rate limited by IP. A key raises the limit and is required only for the tools that are flagged as needing one.
MIT licensed and published on npm as @cresva/acp-mcp.
Claude and ChatGPT reach the same intelligence layer
This page is the MCP door. The other is a Custom GPT. They are two ways into the same thing rather than two products.
Claude
You are here · MCP
Questions
What is MCP?
The Model Context Protocol is an open standard for giving an AI client a set of tools it can call. Claude Desktop speaks it natively. You add a server to a config file, restart, and the tools appear in the conversation.
Does this run locally?
The server runs on your own machine over stdio. It is not a hosted endpoint and there is no Cresva session in the middle. It does reach the network: each tool call is an HTTPS request to api.cresva.ai, which is where the storefront data lives.
Do I need an API key?
No API key is needed to start. Tools run anonymously and are rate limited by IP. A key raises the limit and is required only for the tools that are flagged as needing one.
How do I rotate a key I have set?
The key is read from the CRESVA_API_KEY environment variable in the same config entry. Replace the value and restart Claude Desktop; nothing is cached anywhere else, because the server is a process on your machine that starts fresh each time.
Does it work in anything other than Claude Desktop?
Standard stdio MCP, built on the official SDK, so it works with any MCP-compatible client rather than only with one.
The package is @cresva/acp-mcp on npm. Everything above was checked against the published build on 2026-08-11 rather than against the source in our repo.
Paste the block. Restart Claude.
No account, and no key to paste to start. The tools are there in the next conversation.