See what your agent understood, before it acts.
Type a task in plain words. The waiter repeats your order back: your words on the left, what was understood on the right. Anything you did not say glows, so you can fix it before any work starts.
Underlined was repeated back. Glowing was not used.
A glowing part is something you did not say. Click it to correct it or confirm it.
Order confirmed
ApprovedCopy this into your agent or your ticket. It says what was agreed, with every guess either confirmed or corrected.
How the waiter checks your order
- You order. One task in plain words, the way you would tell an agent.
- It is read back. A model on Cloudflare Workers AI says the goal, the scope, what would be touched, what it assumed and the risk. For every part it must copy the exact words of yours it rests on.
- The words are checked. This page looks for those words in what you wrote. A part with nothing of yours behind it glows, and so does any clause of yours the reading never used.
- You decide. Click a glowing part to correct it or confirm it, and the order is read again. Then approve, edit or cancel. Nothing runs until you approve, and approving only gives you the confirmed order to copy.
Asking an agent to repeat a task back by hand gives you 200 to 270 words to read, with its guesses mixed in among your own words. Here the guesses are the parts that glow.
The bill
Nothing to pay and no account. Each reading is one call to an open model that uses a small part of our free daily Workers AI allowance. Your task is not stored or logged.
- Each reading
- nothing
- Readings from one network
- 30 a day, 6 a minute
- Model
- gemma-4-26b-a4b-it
For agents: read a task back over HTTP
An agent can ask for the same reading before it starts work, and show the glowing parts to its person. Send a POST with a JSON task. Each part of the answer has basis set to stated or filled_in, and order marks each clause of the task as heard or missed. While any part except the risk is a guess, next_step.action is ask_person and next_step.ask_about lists what to ask before acting.
curl -s https://tools.openkrill.app/mirror/read \
-H 'content-type: application/json' \
-d '{"task": "Make the signup page faster on phones."}'To correct a part, send the same task again with notes. Each note names the part by its text from the last reading and says what you meant:
{"task": "Make the signup page faster on phones.",
"notes": [{"part": "Optimise images on the page.", "correction": "Only the form; images are out."}]}Frequently asked questions
What does it mean when a part of the reading glows?
It means you did not say it. The model must copy the exact words each part rests on; the page then looks for those words in your task and your corrections. A part with no words of yours behind it glows, whatever the model claims.
Does Mirror Twin do the task?
No. It only reads the task back. When the reading is right, approve it and copy the confirmed order into your own agent or ticket.
Is what I type stored?
No. Your task goes to a model on Cloudflare Workers AI for one reading and is not stored or logged by us. Only a count of readings per day is kept, to hold the daily limit.
How many readings can I make?
Up to 30 a day from one network, 6 a minute, while the shared daily allowance lasts. A correction is a new reading. The count resets at 00:00 UTC.