( 01 — The brief )
Stop paying frontier prices for everyday work.
Run everything through the most powerful model and you pay top prices for tasks a cheaper one would handle just as well. The client wanted a managed-credits app that fixes exactly that — without asking you to become a routing expert. Top up once, type, and each request goes to the cheapest model that can do it competently, with the exact cost shown every time. It isn’t pitched as charity; it’s honest about the deal — a small cut of what you save. By the company’s own estimate, per-request routing runs 40–75% cheaper than putting everything on a top model, with no noticeable drop in quality.
— The client, in every case on these pages, was us. A demanding client — and, fittingly, a frugal one.
( 02 — The craft )
You never pick a model.
The interface refuses to turn you into a dispatcher: no model menus, no settings. A cheap classifier reads each request once and sends it to the right rung of a cheapest-first ladder — a fast, small model at the bottom, a frontier one at the top. Three steps and you’re running: top up, type, it routes. And it shows its working — every reply carries the model it used and its exact charge, beside a live balance counted to a fraction of a penny. No surprise invoices, ever.
There are two panes on one billing spine: a chat window, and a code pane that runs on your own machine — reading, writing and running your real files inside a folder you choose. The work stays on the computer; the only thing that ever leaves it is the model call itself.
No model menus. No surprise invoices. You never pick; you always know.
( 03 — The engine )
One shared path does all the work.
The cleverness is in how little there is. Every request — from either pane — funnels through one shared function that routes it, calls the provider, meters the real tokens used, adds a small margin, and decrements your balance. Build that billing spine once and everything inherits it. And because the same backend also speaks the industry-standard dialect every AI tool already knows, the identical routing and metering can wrap a coding agent exactly as it wraps the chat — no special integration per feature.
The numbers aren’t guessed: a real price table turns actual token usage into the figure you see, so the charge on each reply is a true cost, not an estimate. Sensible guards sit underneath — bounded output and a daily spend cap — because a prepaid balance wired to an autonomous agent is precisely the sort of thing that needs a seatbelt.
Meter the tokens, add the margin, show the bill — every request, the same path.
( 04 — Foundations )
An early beta that says so.
The honesty runs all the way down to how the product talks about itself. It puts its caveats in plain sight rather than burying them: the browser demo admits it may pause when it reaches a safety limit; the Windows installer is candid that it isn’t code-signed yet and tells you exactly how to get past the warning; and the headline savings figure is presented openly as the company’s own estimate, not a benchmark.
Even the free trial is defended thoughtfully — a dollar of credit to start, no card required — with a quiet check that stops one person harvesting it under a dozen email aliases. Small courtesies, all pointing the same way: tell people the truth, including the price, especially the price.
Honest about the cost. Honest about itself.
( In brief )
- Free to try
- $1, no card
- Choosing a model
- You never do
- Every reply
- Model + exact cost
- Code pane
- Runs on your machine