Correction, 3 August 2026. When this went up in July it said you add ClearList to ChatGPT as a custom connector under Developer mode. That is no longer true for most people, and I have fixed it below rather than quietly editing it out. Custom connectors in ChatGPT are now a Business and Enterprise/Edu feature that a workspace admin has to switch on, so a personal Plus or Pro account cannot add one at all. The route for everyone else is the ChatGPT Plugin Directory, where our listing is still in progress. Claude works on every plan, free included. If you tried to find Developer mode because of this post and could not, that is why, and sorry for the detour.
Almost every company now says AI agents can use their product. Read the fine print and you find the same asterisk: first, go to the website, make an account, find the settings page, generate an API key, and paste it into your assistant.
That is not an agent using your product. That is you doing the hard part and then letting a robot do the typing.
The friction was never the typing.
What we changed
On ClearList, creating the account is itself something the agent can do.
Two of our tools work with no credentials at all. send_verification_code takes an email address. verify_code takes the 6-digit code that lands in the inbox, creates the account if it does not exist, and hands the assistant a key it stores on its own.
The practical result: a person who has never heard of ClearList can end up with a published sale page without a browser being involved at any point.
What the flow actually looks like
You tell your assistant you are moving and need to sell a house full of furniture. From there:
- It asks for your email and calls
send_verification_code. - You check your inbox and read the six digits back into the chat.
- It calls
verify_code, gets a key, and stores it. - You send photos. One at a time, or up to fifty at once.
- It calls
bulk_create_listings. The AI groups photos by item, so four angles of the same dresser become one listing rather than four. Each item gets a title, a description, a price range with the reasoning behind it, estimated dimensions and weight, and a condition assessment. - You look at what came back and correct anything wrong. "The couch is $400, not $250." It calls
edit_listing. - It asks for your city, calls
publish_page, and gives you a link.
You share the link. Buyers browse it with no account and no app, add items to a pickup basket, and verify with their own emailed code. A first-come queue handles who gets what, capped at three people deep per item so nobody waits behind a crowd that will never clear.
Twenty-eight tools sit behind that conversation. You never learn any of their names.
The part I want to be honest about
Reading a code back to your assistant is not elegant. It is a human step wedged into an automated flow, and it exists because the alternative is worse.
The alternative is what everyone else does: make the human open a browser, sign up, navigate to a developer settings page, generate a key, and paste a 68-character string into a chat window. That is five steps of friction to remove one. We kept the one.
There is a browser path too, for what it is worth. If you add ClearList as a connector in Claude, you get a normal OAuth consent screen. If you happen to already be signed in to clearlist.me, you click Approve and you are done. The code only appears if you are signed out, and it goes into our website rather than into the chat.
Two different flows, and which one you get depends on how you connected. Neither involves a password, because ClearList does not have passwords.
Where it deliberately stops
There are 28 tools. One of them takes two calls, on purpose.
Sharing a seller's street address used to be flatly impossible here. We built the tool, removed it, and I wrote a fairly confident post about why it was never coming back.
Then I ran my own sale and found the rule was stranding the exact person it was written for. A seller who sets up everything through ChatGPT never opens the web app, so there is no address saved in it. They could not share through the assistant, and had nothing to share through the app either.
The threat model has not moved. It is someone recognizing a piece of furniture in a public listing, working out whose house it is in, and finding a path to that address that does not require the person living there to make a decision.
What moved is the conclusion. The bar is now that a human is demonstrably the one who shared it, and that the seller finds out every single time. So share_address looks up who would receive the address and discloses nothing, the assistant names that person out loud, and only then does a second call send it. Only a buyer who verified their email by reserving something can receive one, and the seller gets an email afterwards with the address that went out.
The honest caveat: the Approve button is not a security boundary, because the server cannot tell a human tapping it from a model calling the same tool. The longer version of that story is here.
Three other things an agent cannot do here: ship an item, take a buyer's money, or list anything from the prohibited categories. Pickup is in person and payment is between the two people involved. We are not in the middle of that, which is also why there is no commission.
Why this is worth caring about beyond one app
The interesting question is not whether an AI can write a furniture listing. It has been able to do that for a while.
The interesting question is what happens to the shape of software when the person never sees the interface.
Almost everything we know about building products assumes a human is looking at a screen. Onboarding funnels, empty states, tooltips, progressive disclosure, the entire discipline of making the second step obvious once someone has finished the first. All of it assumes eyes.
When the consumer is an agent, most of that stops mattering and different things start to. Can the thing be discovered without a human typing your brand name? Are your capabilities described somewhere a machine can read before it commits to using you? When something fails, does the error explain what to do next, or does it just say 400?
We publish a /llms.txt, an OpenAPI spec, an MCP server card, and agent skill files at /.well-known/agent-skills/. None of that is visible to a person using the site. All of it is load-bearing for an agent deciding whether we can solve their user's problem.
That is the actual shift. Not that AI can write your listing. That the software you use may increasingly be software you never look at.
Trying it
The server is listed in the MCP Registry as io.github.Cuneiform-LLC/clearlist, on Smithery, and on npm as @clearlist/mcp-server. The remote endpoint is https://clearlist.me/api/mcp.
Three items are free, so you can find out whether any of this works before deciding whether it is worth anything to you.
If it does not work, tell me. I would rather hear it from you than not hear it.