Keep every word. Keep it on your Mac.

RecapMuse captures system audio through Core Audio process taps, available on macOS 14.2 and later. It takes a copy of the system output stream, and that is the whole of its reach. It installs no virtual audio driver. It changes no default input or output device. It never reroutes or mutes your audio. It does not join the meeting or appear to other participants. The tap and its private aggregate device exist only in memory and disappear when the app exits, so your Mac's normal audio setup is exactly as you left it.

That detail is becoming load-bearing rather than academic. Microsoft Teams begins automatically blocking identified meeting bots in August 2026, and in September 2026 bot approval becomes a company-wide default for every tenant, so a notetaker that joins as a participant can sit in the lobby waiting for an organiser to admit it. RecapMuse has no participant to admit, because it never joins. What the Teams change actually does, with sources.

The file format is chosen for a recorder that has to survive a crash. RecapMuse writes timestamped Audio Data Transport Stream ("ADTS") Advanced Audio Coding ("AAC") files, one an hour by default, into a folder you pick. Each file grows continuously on disk and stays fully playable if the process is force-killed or the Mac reboots mid-file. There is no finalize step to miss. The files open in QuickTime, VLC, Music, and ffmpeg with no conversion.

At 96 kbps, which is enough for voice, an hour is about 43 MB and around-the-clock recording runs to roughly one gigabyte a day. You choose where the files land, how long each one is, what bitrate they use, and whether the microphone is included at all. Pointing the folder at a synced drive is your call and your bandwidth.

Out of the box it records system audio only, which needs no microphone permission at all. Turning the microphone on asks for Microphone permission once, and the two modes then behave differently in a way worth knowing before you choose: with the microphone off, the tap produces samples only while audio is actually playing, so stretches of pure silence collapse rather than being recorded as silence. With the microphone on, the microphone holds a continuous timeline.

And one honest detail about knowing it is on. File size is not the signal, because an hourly file grows in bursts and a silent stretch with the microphone off grows barely at all. The status check reports whether the tap itself is running, which is the thing you actually want to know, and it lists the recent files alongside it. A recorder that leaves you guessing whether it caught the meeting has failed at the only job it has.

One more thing that is easier to say now than to discover later. Starting the app inside your normal login session inherits audio-capture permission from whatever launched it. Running it across reboots makes it its own process, so macOS grants it a separate Audio Recording permission the first time. That grant is tied to the specific binary, so rebuilding or moving it can make macOS revoke it and ask again.

Version 1.8.0 replaced the single blanket privacy switch with four independent ones in Settings: transcription, diary summaries, calendar auto-naming (suggests a recording's name from an overlapping calendar event, read on-device, you confirm before anything renames), and cloud sync. Each is set on its own, because they are genuinely different decisions. Ask is separate from all four: it always runs on-device, with no cloud option and nothing to switch.

Transcription and summarization run on separate engines, which is what makes the split real rather than cosmetic. Keep transcription on your Mac while allowing a cloud summary, keep summarization local while sending transcription out, run both on-device, or send both. A "Maximum privacy" master toggle moves every feature to its most private state in one tap, and the Session edition, when it ships, will set both engines to on-device by default.

Each of the four toggles carries a live sentence stating whether OpenAI is involved in that feature, and where it is, the sentence confirms that OpenAI does not train on API data. The state is on the screen rather than in a policy document.

Ask puts the evidence first. Query your past conversations and a "Sources (N)" card renders above the answer rather than as chips underneath it, so you can look at what the answer is built from before you read the answer. Every source links straight into the conversation it came from, and past eight sources the card paginates. An answer you can check is worth more than an answer you have to trust.

Recording other people may require their consent, depending on your jurisdiction. That is worth knowing before you switch the microphone on, and it is not legal advice.