The difference
Atrytone, or just asking an AI assistant?
An AI assistant in your editor or terminal changes your code when you ask it to. Atrytone is the process around that kind of work: it finds what to ask for, makes the change on a copy, waits for your yes, looks again after every change, and every night, and keeps a record. If you review every change yourself and need no record, an assistant alone can be enough.
Last updated
What an assistant alone gives you
You decide what to ask, the assistant changes the code you point it at, and you review what it did. For a small change that you will read yourself, that is a fine way to work.
What Atrytone adds
- A list of what is missing. We read your code for 23 of them today, as rules and not as an AI’s opinion. For Tests, we run your app on a copy. We do not test your live site. Nothing changes without your yes.
- A copy and your yes. Every fix is made on a copy and waits for your yes: “Yes, merge it into my app”. Atrytone merges the fix into your GitHub repository on your yes. It does not deploy: if your host publishes from GitHub, the merge goes live there.
- A second look. It reads the code again after every change, and every night.
- A record. Who asked, which model made the change, what was checked, what it cost and who said yes. Each entry carries the fingerprint of the one before it, so a changed entry shows, and our app can only add to the record. Anyone can check a receipt themselves.
- Your own models. Every fix job so far has run on Claude. OpenAI and Gemini keys connect today; no fix job has run on them yet. Your model company bills you separately for the fixes, on your own key.
When an assistant alone is enough
- You are changing one thing, and you will read the change yourself.
- You already have tests and a habit of reviewing, and nobody needs a record of who said yes.
- The app is a prototype nobody depends on.
When Atrytone is worth it
On 10 November 2026, 8 fix jobs had been asked for from the check's results, all in our own workspace and none in a customer's. Our own bar before we call fix jobs reliable: five in a row, on apps we did not hand-pick, each reaching its goal with nobody stepping in. On 10 November 2026 we were at 0 of 5.
- Real customers will use the app, and you do not know what a real product needs that yours may lack.
- Someone else must be able to see what changed, who said yes and what was checked.
- You want the app looked at again every night, not only when you remember.