MCP

Shop from your terminal.

Most of what is sold here is a Claude Code artifact — a skill, a pipeline, a prompt set. Bought on a web page it is a zip in your downloads folder. Bought here it is installed, and the next thing you type can use it.

Connect Browse first

What it can do

`search` · `browse` · `product`	Find things, and read the structured spec — contents, what you need already, and what it explicitly does not do.
`product_image`	Look at it.
`start_checkout`	Returns a payment link for you to open. No tool here can spend your money.
`install`	Unpacks a purchase into the directory you name.
`listing_template` · `publish`	Selling: point it at your files, let your own Claude draft the listing from what is actually in them, and publish.

Connect

The hosted endpoint is not up yet. This shop runs on one machine while it is being finished, so there is no public URL to point at and it would be dishonest to print one here.

Running your own copy in the meantime:

claude mcp add nabemall -- python /app/mcp_server.py

Browsing and buying need no account. Selling needs a key — generate one in Settings, or call issue_api_key once with your sign-in details.

Two things worth knowing before you trust it

It will not tell you what is good. We wrote most of this catalogue, so a ranking authored by this server would be an advertisement in a lab coat. Every result carries price, contents, file list, sales count and a house flag saying whether we made it. Your own model does the judging, which is the right way round.

Listing text is written by strangers. It arrives under keys named untrusted_ for exactly that reason. A product description that says "ignore previous instructions" is a description anyone can write for free, and naming the field is a stronger defence than a warning that can be summarised away.