You have a room full of things you have been meaning to sell, and you have Claude open in another tab. Those two facts are more connected than they were six months ago.
This is the practical version of that. What to click, what happens next, and where it stops.
Setting it up
ClearList runs a remote MCP server at https://clearlist.me/api/mcp. In Claude, you add that URL as a connector in settings, and the tools show up in your next conversation.
That is the whole setup. This post is about using it in the Claude app, which is what most people want. If you live in a terminal and would rather point something at a folder of photos and walk away, that is a different workflow and it has its own post.
What signing in actually looks like
Claude runs a standard OAuth flow when you add the connector. You get a consent screen, and if you are not signed in already, a login first.
That login is passwordless. You put in your email, ClearList sends a 6-digit code, you type it in. No password to invent, no settings page, no API key to copy and paste somewhere with a warning label about keeping it secret.
What you never do is the thing most "works with AI agents" products still require: go to the website, make an account, find the developer tab, generate a key, come back and paste it. That is us doing the tedious part and then letting the robot do the typing.
What the conversation actually looks like
Twenty-six tools sit behind the connector, though you will never think about them as tools. Twenty-three are the seller actions. Three are stubs for a search feature that does not exist yet, and they say so when called, because a tool that quietly does nothing is worse than one that admits it. There is no signup tool here, because the OAuth flow above already did that job; the npm package carries two extra tools for signing up in conversation, which is how you would do it in a client with no browser.
In practice it goes like this.
You send photos. Claude comes back with titles, descriptions, prices, dimensions, weight, condition, and whether each thing needs a truck to move. You say "the dresser is more like $80, and the lamp is not for sale." Both get fixed. You say "publish it," and you get a link.
Buyers land on that link and reserve things. A maximum of three people can queue for any single item, so you are not fielding fifteen messages about one bookshelf. Buyers verify an email and nothing else. No account, no app download, no "is this still available."
Then you can ask Claude who is waiting on what, and reply to them, without opening anything.
Where it earns its keep
For three items, honestly, it does not. Open the website, spend four minutes, done.
The chat version starts winning somewhere around item eight. Past thirty or so, uploading photos a handful at a time becomes its own chore, and at that point you want the folder workflow instead.
It is worth being precise about why chat is better in its range, because it is not that the AI is smarter in a chat window. It is the same engine either way. It wins because you never leave the place you were already working, and because "make all the kitchen stuff free, I just want it gone" is a sentence that does something.
Same product. It just meets us where we already are.
The one thing it will not do alone
Claude can share your address. It cannot decide to.
Ask it to, and nothing goes anywhere yet. It comes back with a card naming the person: Sarah M., her email, and the exact address underneath. You tap send. Then you get an email recording that it went out, and what went out, which no assistant can switch off.
The threat model is not hypothetical. Someone who wants to find a person can look at a photo of a living room, recognize the couch, and message the seller as an interested buyer. It looks exactly like a sale right up until it is not. All of the caution in that moment belongs to whoever lives there.
For a while I thought the only safe version was not building it, and I wrote that here. Then I ran a sale on my own product and found the rule was stranding people who had set everything up through Claude and had no address saved in the app to share. So the rule changed shape: a person has to be the one who says yes, and they have to find out every single time. Only a buyer who verified their email by reserving something can receive one at all.
And the honest caveat, since it is the part nobody prints: the tap-to-approve button is not a security boundary. My server sees the same key whether you tapped it or the model did. The longer version of that story is here.
What to try first
Take one photo of one thing you have been meaning to sell for six months. Send it to Claude. See what comes back.
If the price looks wrong, argue with it. It will adjust, and watching it adjust will tell you more about whether you trust the thing than any guide I could write.
That room full of stuff in the other tab is still there. This just moves it out in an afternoon instead of a series of weekends.
Related reading: Can ChatGPT sell my stuff? and how to sell a ton of stuff without writing separate listings.