Private AI Notetaker for Mac: The Audio Path Decides
By Kyle Nelson, Founder, Fazit
A private AI notetaker for Mac is not a marketing label. It is a pipeline. The question that matters for confidential client calls is whether the tool creates an audio file, sends it to a vendor, or keeps the sound in RAM until text exists. Fazit is a native macOS menubar app that turns 1:1 client calls into finished Markdown notes in an Obsidian vault, with no bot joining the call, transcription and note generation running on the Mac, and call audio never written to disk.
What "Private" Actually Means on a Mac
The word private gets stretched thin in this category. A tool can be bot-free and still stream your audio to a vendor cloud. A tool can be local-first and still write an audio file to your disk. A tool can be on-device and still keep a transcript in a vendor database. The only definition that survives a security review is architectural: where in the pipeline a durable artifact of your conversation is created, and who holds it afterward.
Fazit's pipeline is specific. A Core Audio process tap reads the call app's output and the microphone directly. Nothing joins the meeting and no other participant sees anything. Audio exists only in a fixed-size RAM ring buffer with no write API, destroyed on every exit path including errors. Every note records audio_retained: false in its frontmatter. Transcription runs on-device using Parakeet via CoreML, and notes are generated by a local model on localhost. The output is a plain Markdown file written into a folder of your Obsidian vault, or Apple Notes. There is no Fazit database holding a copy.
This is not the same as "bot-free." Bot-free is not the same as private. Granola, Krisp, and Circleback all capture meetings without a bot, but their audio still streams to vendor clouds and third-party speech services. They removed the bot. They did not remove the server.
Local Transcription on Apple Silicon
Speech recognition moved to the cloud a decade ago because laptops could not run the models. That changed. Apple Silicon's Neural Engine and NVIDIA's Parakeet TDT v3 ended the workaround: a server-class speech model now transcribes your calls on a MacBook, faster than the conversation happens. The engineering story is covered in Cloud Transcription Was a Hardware Workaround. The Hardware Caught Up.
The practical consequence for a consultant or lawyer is that you no longer have to choose between a good note and a confidential call. Fazit captures two audio streams separately: your microphone ("You") and the call app's own audio ("Them"). Transcribing them separately gives perfect You/Them attribution on a 1:1 call with no diarization guesswork. The audio is transcribed in memory and never written to disk.
Several Mac tools now claim local transcription. Private Notetaker says its meeting notes, transcripts, and action items are generated locally on the user's computer, with support for Apple Silicon Macs, according to its site. Weeve states that recording, transcription, and summaries run on the user's device and can work online or offline. Kimiko supports Apple Silicon Macs running macOS 14.4 or later and provides local transcription while storing notes on the user's computer. Resonant says its Mac app captures system and microphone audio directly on the computer and transcribes locally on Apple Silicon.
The difference is what happens to the audio before transcription. Local transcription is necessary but not sufficient. A tool can transcribe locally and still write an audio file to disk first. A tool can transcribe locally and still upload the transcript to a vendor for summarization. The question is whether the tool creates an audio file, sends it to a vendor, or keeps the sound in RAM until text exists.
The Audio File Is the Liability
The Otter.ai wiretap class action gave the AI notetaker legal risk a case number and a price: $5,000 per recording under California Penal Code § 637.2, plus a transcript that becomes discoverable evidence in any lawsuit. The case is covered in Brewer v. Otter.ai: The Case File, and Where It Stands. One February 2025 sales call, captured by a notetaker the plaintiff never installed, became four consolidated class actions. No court has ruled yet, but the complaint is a checklist of what any cloud notetaker does with your calls.
The legal exposure points at the same artifact: the recording. Discovery and wiretap statutes both target the audio file. The one you never made can't be subpoenaed. This is not legal advice, and consent and professional-conduct duties still apply to the person running the call. But the architecture removes the vendor-as-interceptor theory and the audio-file exposure.
For lawyers, the privilege question starts with where the client-call audio, transcript, prompts, logs, and summaries go. Sending client-call audio to a vendor and its subprocessors is a third-party disclosure that invites waiver arguments and creates a discoverable record. The analysis is in AI Notetakers for Lawyers: The Attorney-Client Privilege Risk of Cloud Transcription. A local workflow reduces third-party disclosure because call content stays on the Mac and call audio is never written to disk.
For consultants, the NDA you signed already decided the question. Every consulting engagement starts with a confidentiality clause: no disclosure of client information to third parties without written consent. A cloud AI notetaker on a client call is exactly that disclosure. The tool that satisfies the clause is the one that never sends the audio anywhere, as covered in An AI Notetaker for Consultants: The NDA You Signed Already Decided.
The Comparison That Reorders Around One Question
Five tools everyone shortlists, and they all write a good note now. Fathom, Otter, Fireflies, tl;dv, and Granola each win a different job: free tier, real-time archive, sales CRM, video, bot-free notes. Rank them on the only column that survives a security review, who holds a copy of your conversation after the call ends, and the list reorders. Four keep a recording. One deletes after transcription. Fazit is the sixth row, where the copy was never eligible to exist. The full comparison is in Fathom vs Fireflies vs Otter vs tl;dv vs Granola: The Comparison Reorders Around One Question.
The table below is the short version for a Mac user with confidentiality obligations.
| Tool | Bot joins call | Audio leaves Mac | Audio file created | Note location |
|---|---|---|---|---|
| Fathom | Yes | Yes | Yes | Vendor cloud |
| Otter | Yes | Yes | Yes | Vendor cloud |
| Fireflies | Yes | Yes | Yes | Vendor cloud |
| tl;dv | Yes | Yes | Yes | Vendor cloud |
| Granola | No | Yes | Yes | Vendor cloud |
| Fazit | No | No | No | Your Obsidian vault |
Granola is the best of the bot-free notetakers, but it still streams your audio to third-party transcription services. Its public security FAQ names OpenAI and Anthropic, uses US-only AWS storage, and retains transcripts indefinitely by default. The details are in Granola Subprocessors: What the Public Docs Actually Say.
What a Finished Note Looks Like
The output is a plain Markdown file written into a folder of your Obsidian vault. There is no Fazit database holding a copy. The note captures the client, date, objective, decisions, commitments, owners, due dates, and a paste-ready follow-up email in the client's own language and register. The email draft is the moment people screenshot.
A meeting note is only yours if it is a file you can grep. Cloud notetakers store your notes as rows in a vendor's database, which means you can't search them with your own tools, you can't back them up with your own system, and you can't delete them by deleting a file. The argument is in Obsidian Meeting Notes Should Be Files, Not Rows in a Vendor's Database.
Fazit writes one Markdown file per call into <vault>/Fazit/. The frontmatter records audio_retained: false. The body contains a clean summary, key moments, action items with owners, and a You/Them transcript. No audio anywhere.
The Consent Question Remains Yours
Fazit's architecture reduces the vendor's and the audio-file exposure. It does not remove your duty to have consent where the law requires it. California CIPA is all-party. Germany's §201 StGB is criminal. The consent question applies to every bot-free notetaker, including Fazit, as covered in The Granola Lawsuit: What Chamberlain v. Granola Alleges, and What It Does Not.
The honest disclosure is simple: audio really isn't kept, and the other person gets a real chance to say no. That is a different conversation from explaining why a bot is in the room and where the recording goes. In a documented 6-week test of 47 meetings, 32% of clients questioned a visible recording bot when Fireflies or Otter joined, versus zero questions for bot-less system-audio capture. Invisible capture is worth paying for.