Coral went up on Hacker News this week as a Show HN. It's an open-source, self-hosted gateway that sits between AI agents and the systems they work on. Agents write short programs to call tools, then save the ones that worked so tomorrow's run doesn't start from zero.
The first of the top comments in the thread went somewhere else, and that's the part I keep coming back to:
What would be great too is a library for mini apps built with agents. I find these little apps to be useful all the time, but despite usually being written in very versatile HTML that is cross-platform, these things are confined to the specific platform I used and usually the chats too.
It's a real problem, and you don't have to be a developer to run into it.
The app that lives in one chat
Ask a chatbot for a tuition tracker, a shift sheet or a little tool that adds up today's orders, and you'll usually get something nice that works. Then:
- It lives in that one conversation. Next month you're scrolling up to find it, or asking for it again.
- It forgets. Close the tab and whatever you typed in is gone, because there's nowhere to keep it.
- Nobody else can use it. Your staff can't open your chat, and with no sign-in it's either everyone sees everything or nobody uses it.
- It only runs while you're watching. It can't send tomorrow's bookings at 8 PM.
Coral's answer for agents is to save the code that worked. The commenter wants the same for the apps. I'd go a step further and give the app a home outside the chat altogether.
What a mini app needs to move out
Four things, mostly boring:
- An address. A link you can bookmark and send to someone.
- Somewhere to keep data. So what went in on Monday is still there on Friday.
- Accounts and roles. So staff sign in and see their part.
- A schedule. So it keeps working after you close.
That's what DontCode sets up before you ask. Every project starts with a database, sign-in, email and hosting, and every change Insight builds goes live on its own. You describe the tracker in a chat like you would anywhere, and what comes back is an app with its own link.
Limit what it can do
There's another idea in Coral's write-up worth borrowing. Most gateways only decide which tools an agent can see, and Coral calls that all or nothing: if an agent can call update_record, it can change any record the tool allows. Their fix is a narrow tool like closeStaleTicket, which changes one field, only on tickets with no activity for 30 days.
A small business needs the same thing for people. The staff-only pages recipe gives each staff member a sign-in where they see today's orders and bookings and mark them ready, while sales, refunds and settings stay with you. Behind that, the app's pages can only reach the tables they actually use. On a plan with checks, Insight also writes checks that run as staff, covering what they can do and what they shouldn't, and a build has to pass them before it goes live.
A quick test
Before you rebuild the same chat tool a third time, ask three things. Does it need to remember anything? Does anyone else need to use it? Does it need to run when you're not around? If all three are no, the chat is a fine place for it. If any one is yes, it's an app, and it should get to live like one.
If one of those tools is sitting in your chat history right now, try describing it to Insight on DontCode. It's free to start.