What happens on a call, and after it.

When a meeting is detected, a small card appears: which app is using the microphone, and three choices: Record, Not this one, or Snooze, which defaults to 10 minutes. Until you click Record, no audio is written anywhere.

If you ignore the prompt, it goes away on its own. A recording starts without the card only for an app you put on the automatic-recording list in Settings (the list is empty until you add one), or from a connected agent if you turn on Allow agents to start recordings in Settings → Integrations yourself.

Telling the people on the call is on you, and the card gives you the words: Copy notice. The moment you gave the recording notice is kept with the recording too, and a legal hold stops anything about that call being deleted, unless an organisation has set a retention limit.

Turn it on under Settings → Record prompt. Once the transcript is ready, the passport sits under it: how the notice reached the others (copied, or shown to read out) and when you marked it as given; whether a plain yes in English was heard from the other side in the first minutes, with the time and the line it points at; and a one-line note about the rule where your machine's region is set, linking the recording-law guide as general information, not legal advice.

It is conservative on purpose. A hedge, a condition, a question, another language or a reply to something else is "not detected", and a line that may be an objection is pointed out. It reads the words as first transcribed, so an edit cannot make a yes appear, and it never says agreement was given: Recordist cannot know who else was on the call or what they heard, and the audio is the record.

When the recording stops, the notes are written: a summary, the decisions, the open questions and the action items, each ending with the time it was said. On Pro, the follow-up email comes with them, with nothing to set: a short recap and the next steps, from you to the others on the call, for you to read, change and send from your own mail. Recordist sends nothing. On the Fast built-in model, no email is filled in for you. You tick the lines you want, and the draft is made from exactly those. The built-in engine writes the notes on your machine, with nothing to install and no account or key, and the transcript stays there. If you would rather use a frontier model, paste your own API key: Recordist calls that provider directly and never proxies through a Recordist server.

Before the notes are saved, each decision, action item and open question is checked against the line of the transcript it comes from, and each quotation against the transcript's own words. An item the transcript does not support is removed, an owner or due date it does not state outright is labelled as worked out, so you know to check it, and the meeting says how many items were removed or changed. Each decision, action item and open question ends with the time it was said, which scrolls the transcript to that line. The check compares words, so notes can still be wrong: check anything that matters against the transcript.

Ask (Pro) puts a question to your last 5 or 8 meetings, or to the meetings your search found ("what did Ana say about the deadline?"), and the answer names the meeting and the time it comes from. Templates control the shape of the notes: use the built-ins or write your own.

Promises gathers the commitments in your notes into two lists, You owe and You're owed, each with the words it was made in and the time; while the audio is kept, the line plays in the voice that said it. Filter by person to see what stands between you and one client. Tick one done and Recordist offers a short note to tell the other person; a promise made to you that is due offers a gentle reminder. Both open as a draft in your own mail, and Recordist sends nothing.

After a call, once no other app is using your microphone, a small recap card shows the first decision and the first promise each way, and the menu bar keeps one line for the day: the calls, the time, and the promises you owe that come due.

The Promise Ledger, under Settings → AI notes, is off until you turn it on. When a later call says a promise of yours was kept ("I sent you the revised scope"), Promises offers Done with the words that said it, and you confirm; nothing closes by itself, and only a promise whose line the transcript holds and whose owner is you counts. Remind me the day before sends one notification on this machine. The Integrity Ledger shows your own promises this quarter, made, kept, kept late and still open past their date, with a percentage only once ten are decided. It is shown only to you, and nothing of it reaches the local API, the MCP gateway or an export.

Type a word and see the meetings it was said in, with the sentence. Click a result to open the meeting at that line, then click the time to hear it while the audio is kept.

Star the meetings that matter: Delete meetings after never removes a starred meeting.

Whisper runs on your machine, with the speech model you pick kept in a folder you control. Transcription starts while the meeting is still running, so the transcript builds while you talk.

Language is detected automatically across 99 languages, or pinned to one.

Turn it on under Settings → Audio. During a call, press ⇧⌘L, or use the menu bar or the recording strip, and the last 15 seconds are replayed to you; press again to stop. Choose anything from 5 to 60 seconds. The replay comes from the recording itself, so nothing is lost from the file and nothing is added to it.

Headphones keep the replay to you. With speakers instead, your microphone picks it up, and the setting says so. Instant Replay is on the Mac only.

The app writes its network requests to a ledger you can open at any time in Settings → Privacy, apart from a few kinds the help article lists: when, to which host, how many bytes, why, and where that host is (this machine, the local network or the internet). Each request that can carry meeting content is written before it is sent, so one that fails or is cut off still has its row. With the default local setup it lists the downloads the app needs, the update checks at launch and every six hours, and the app's requests to the notes engine that runs beside it, marked as staying on your machine, because audio is captured, transcribed and summarised there.

If you choose a cloud provider for notes with your own key, the ledger records each request that carries a transcript, and to whom. If you run your own Ollama somewhere else, those requests are marked Local network or Internet. No notes request carries audio.

Open it from Settings → Privacy → Privacy receipt. Every figure is a count or a sum over the database and the Privacy Ledger on your machine. Nothing is estimated, and making a receipt sends nothing anywhere.

It shows 0 bytes of meeting content only when the ledger covers the whole period and records no meeting content leaving the machine. When you use a cloud key, it lists how many bytes went to which host. You can copy the receipt as text or save it as an image, and an optional field takes a monthly price you type and shows what it comes to over the period.

When an event in your calendar, from your Mac's calendar or a calendar feed, with attendees starts within 45 minutes, Recordist matches the attendees' names against your past meetings and shows a short brief on the Meetings screen: a summary of the last meeting, the decisions and the action items still open, which you can tick off there.

The brief is put together on your machine from your past meetings' notes and action items, with no AI model. It needs no account and no network, and it is part of Pro and of the 7 days of Pro every install starts with. On Windows and Linux it reads the calendar feed you add under Settings → Detection.

Recordist sits in your menu bar and watches for the signals a meeting gives off: your microphone switching on, the app that holds it, a calendar event starting, and the meeting app's window title.

Detection is entirely local. Recordist reads which app holds the microphone and, if you allow it, the meeting app's window title. It never records your screen and it never joins the call.

Recordist records your microphone and the meeting's audio separately, so what your microphone hears is labelled You. When the microphone also picks up the other side, the copy heard on both channels is kept on theirs.

Everyone on the other side of the call shares one label. When it is still "Speaker" after the call, Who was this? offers the meeting's attendees from the invite: one tap names every line under the label and gives the meeting's promises to that name. With two or more attendees the row asks first, and Undo puts it all back. When the last meeting of the same series named the other side, the name comes first. Names come from the invite or your tap, never from a voice. Telling the other side's voices apart within one meeting is planned; it will not recognise anyone across meetings.

One configuration block and your assistant can list, search and read your meetings, and answer with the timestamp. Ask Claude what was decided in Tuesday's call and it quotes the transcript.

For agent frameworks, Recordist also presents itself as an A2A agent. Both connections exist only on your own machine; no port is opened to your network.

A meeting exports as one Markdown file named with its date and title. At the top are properties that Obsidian shows as such: the title, date, start time, length, attendee names, tags and template. Then come the reminder to check against the recording, the summary, decisions, open questions, the action items as a checklist, and the transcript with a time and a speaker for each turn.

Export to folder saves the file wherever you pick, such as an Obsidian vault. Auto-export Markdown after each meeting, off by default, writes the file to your export folder once the notes are written. Recordist never replaces a file that is already there, so a copy you edited in your vault stays as it is. It has no Notion connection and sends nothing anywhere to export, and Notion imports the files through Import → Text & Markdown. Text, SRT, VTT and JSON are there too, for plain text, subtitles and your own scripts.

A macOS configuration profile can set fifteen managed keys: notes on the Mac only, no cloud AI keys, a cap on how long audio and meetings are kept, no audio export, no agent-started recordings, no automatic recording, no automatic export, one folder for all exports, the organisation name shown next to each managed control, a legal hold, no update checks, and plain http to an AI server on another host, and from 1.0 the Promise Ledger, Instant Replay and the Consent Passport kept off.

Recordist applies the policy on each Mac and locks each managed control with a "Managed by" note. There is no Recordist server or admin console, and the app reports nothing back, so compliance is checked on each Mac. The managed settings page has a sample profile, a schema for Jamf Pro and the list of what isn't managed yet.