What did you just agree to?

Any export where each line starts with the speaker's name: Zoom, Google Meet, Otter, Granola, Fathom, Fireflies, or notes you typed yourself in the form “Sarah: I'll send it Thursday”. Timestamps are fine and are ignored. A block of text with no names cannot work, because everything depends on knowing who said what, and the page tells you that the moment you paste.

About 400 lines in one run, which is roughly a two-hour meeting, and up to 120,000 characters. If yours is longer, run it in parts. Long transcripts are read in parallel chunks, so a long meeting does not take proportionally longer than a short one.

It splits the list into what you owe and what you are owed, and puts your promises first in the copied checklist. You can change the name after the results appear and the list re-sorts straight away. It does not change what was found.

Yes. “Copy checklist” puts the visible promises on your clipboard as a checklist grouped by person, and each card has a copy button for its own line. There is no account, so nothing is saved for later: copy what you need before you close the tab.

Four likely reasons. The line was phrased as a question or a suggestion, so it was not an obligation. The speaker's name was not on it. It fell past the 400-line limit. Or it was a restatement: when the same speaker repeats a promise a moment later (“I know. I'll do it.”), the two are shown as one card, keeping the longer line and the higher risk of the pair.

Each line is put to a decision model as three plain questions: is the speaker taking on an obligation, how firm is it, and how likely is it to be quietly forgotten. Questions, opinions and suggestions nobody accepted are not obligations, so they are dropped. An acknowledged debt, like “I owe him an intro, I keep forgetting”, counts, and is usually the most fragile one in the room.

It reads the name off the line. Every meeting tool writes the speaker into its transcript, so attribution is parsing, not guesswork. That means it cannot get confused and tell you that you owe something you never said. The model is only asked the parts that genuinely need judgement.

They are the evidence, read straight from the sentence and shown so you can check the call in a second. A plain underline marks a deadline, amber marks a hedge such as “probably”, and red marks an admission such as “I keep forgetting”. They sit beside the risk band as evidence you can check. They are not what produced it.

It is a judgement about how the promise was made, not a prediction of what happens next. A commitment with a named date, a clear owner and no hedging reads as solid. One with no date, or one where the speaker admits they have already forgotten it once, reads as fragile. It is shown as a band rather than a percentage because there is no outcome data behind it, and a number like 68% would imply a precision it has not earned.

Because the promise's own sentence names no day or date, and the tool will not borrow one from a nearby line. If it did, it would sometimes put the wrong date on your promise, and a wrong deadline does more harm than an empty one.

Jev is a decision model from TypeSafe, reached through OpenRouter. Instead of writing text, it answers questions put to it about a piece of text with a probability, a choice or a score. That is why nothing on a result is generated prose: the only model output is three answers per line.

We have not scored it against a labelled set of real meetings, so we will not quote an accuracy percentage. What we have tested is repeatability. In September 2026 we ran one 59-line meeting through it five times and got the same 15 promises every time, with risk scores that moved by at most 0.06 on a scale of 0 to 1. Treat it as a fast first pass to check against what you remember, not as a record of the meeting.

No. Speaker names and line breaks are worked out in your browser. When you press the button the lines go to our server, which holds them in memory for that one request and sends them on to the model to answer the three questions per line. Nothing is written to our database. The one thing we do keep is a request counter tied to your IP address, which limits runs to 12 an hour and holds no content.

It goes from our server to OpenRouter, the service that routes the request to TypeSafe's Jev model. We do not control how those providers handle data. If a meeting is sensitive, edit out anything you would not send to a third-party service before you paste it.

No. Results exist only in your browser tab, and there is no shareable link. “Copy a line to share” copies one sentence with the number of promises found, not the transcript or anyone's name.

It is one half of it. This page finds the promise. The product is what makes sure it happens: Mailient watches the threads those commitments live in and drafts the follow-up when one starts going quiet. The tool is that idea shrunk to something you can test in a minute without connecting anything.