Obsidian meeting notes should be files, not rows in a vendor’s database.

A meeting note is only yours if it exists as a file you can open, grep, sync, and back up without asking anyone’s permission. Everything else is a lease.

Three ways to get AI meeting notes into Obsidian

Obsidian ships no AI of its own, so every setup is one of three shapes. They differ in where the audio goes, where the summary is generated, and what actually ends up in your vault. All three produce Markdown, which is why the differences are easy to miss.

1. A community plugin running models you install yourself. Plugins such as AI Meeting Notes capture mic plus system audio, shell out to a Whisper CLI you installed, and pipe the transcript through a local model like ollama run llama3. Fully local, free, and honest about it. The cost is setup and upkeep: you install and maintain the binaries yourself, the README notes that not every OS and hardware combination exposes system audio capture, and the plugin stores the audio files in your vault alongside the notes.

2. A bot-free Mac app that writes into a folder in your vault. Shadow is the clearest current example: it detects meetings at the system level with no bot, transcribes on-device on Apple Silicon, and drops Markdown into a folder you can point at your vault. Read the terms carefully, because bot-free and on-device are two different promises. Its pricing page lists unlimited audio recording as a feature, so the recordings are kept, and its privacy policy says that when you use the AI features, certain content including transcripts may be shared with external servers, naming OpenAI under no-training terms. Local capture and on-device transcription, cloud-assisted notes. Free tier, $8/month for the AI notes. (Both checked 1 August 2026; vendor terms move, so re-read them yourself.)

3. A cloud notetaker plus a sync bridge. Granola, Otter, Fireflies and the rest process everything on their servers, and a community plugin copies the result down into your vault. It works, and we have written the guide for doing it with Granola. The structural catch is that the vault copy is downstream of a vendor database that stays authoritative, and the bridge is now load-bearing infrastructure you do not own.

Where Fazit sits. Shape two, with the two remaining cloud dependencies removed. Capture is bot-free, transcription is on-device, and the summary is generated by a local model on localhost rather than a provider API, so no transcript leaves the Mac either. The audio is never written to disk at all, which is a stronger claim than deleting it afterwards. The only network traffic is the one-time model download from public CDNs. If you are weighing these tradeoffs generally, the longer version is in on-device vs cloud AI notetakers, and the reason bot-free alone is not a privacy claim is in bot-free is not private.

The SaaS notetaker data model

Every cloud meeting assistant stores your conversations the same way: rows in a multi-tenant database, behind an account, behind a subscription, behind an API with rate limits. The “export” button is an afterthought — usually a ZIP of HTML or a PDF that strips the structure you’d want to query.

That model has three failure modes, and none of them are hypothetical:

• The vendor changes terms. Retention policies, training carve-outs, and pricing tiers get rewritten under you. Your two years of client history are the leverage.
• The vendor gets breached or subpoenaed. A centralized store of thousands of companies’ conversations is exactly the target the last decade of breach disclosures describes.
• The vendor shuts down. Meeting-tool consolidation is ongoing; sunset emails give you 90 days and an export queue.

If your client work has any confidentiality dimension — coaching, consulting, legal, finance — each of these is not an inconvenience. It is a conversation you owe a client.

What a note-as-a-file looks like

Fazit’s entire output is a Markdown file written into a Fazit/ folder inside your existing Obsidian vault. Not a plugin database, not a sidecar format — a file named 2026-07-16 1030 Client kickoff.md, written atomically, that Obsidian indexes like anything else you wrote by hand.

Each note opens with plain YAML frontmatter:

---
source: Fazit
started: 2026-07-16T09:30:12Z
ended: 2026-07-16T10:02:47Z
audio_retained: false
---

Below it: the summary, key moments, action items with owners, a paste-ready follow-up email, and the full You/Them transcript. When the call ends, Fazit opens the note directly via Obsidian’s obsidian://open?path= URL scheme — the note appears in your vault seconds after you click Stop.

Frontmatter is an audit surface

Because the metadata is structured YAML rather than a vendor dashboard, your vault becomes queryable with the tools you already run. A Dataview table of every client session this quarter:

TABLE started, ended
FROM "Fazit"
WHERE audio_retained = false
SORT started DESC

Note the WHERE clause. audio_retained: false is stamped into every note not as a compliance checkbox but as a description of the execution path: audio lives in a fixed-size ring buffer in RAM, transcription consumes a snapshot, and the buffer is reset unconditionally on every exit path. The note can truthfully carry that line because there is no code path that writes audio to disk. The full argument is in Why “Never Records” Is Not Marketing — It Is an Invariant.

That means an auditor — or a skeptical client — asking “where are the recordings of our calls?” gets a one-line answer: there never were any. The distinction between never created and created then deleted is the entire game, and it is visible in every file’s first ten lines.

The pipeline that makes local-only possible

None of this requires a server because the whole pipeline runs on Apple Silicon:

• Capture: a process tap on the call app plus the mic, into RAM.
• Transcription: on-device ASR — no audio leaves the machine.
• Note generation: a local model produces the summary and follow-up email; the transcript never touches an API.
• Write: one atomic Markdown write into your vault. That write is the only persistence in the system.

In 2024 this stack required a gaming GPU and patience. In 2026 it runs quietly in a menubar app on the laptop you already own. The economic argument for shipping your client conversations to a datacenter is gone; what remains is vendor lock-in dressed as convenience.

Data sovereignty is now a client-facing question

Regulated and privacy-sensitive professionals are increasingly being asked by their clients what tools touch the conversation. EU clients ask about transfer mechanisms; enterprise procurement asks for subprocessor lists. A cloud notetaker adds a subprocessor. A local pipeline that writes plain files into a vault you control adds nothing to the list — there is no third party to disclose, because there is no third party.

(As always: this describes architecture, not legal advice. Consent for transcription is on you, and Fazit ships a one-line disclosure notice for exactly that.)

Plain text also solves the boring half of sovereignty: portability. Markdown files sync with Obsidian Sync, iCloud, Syncthing, or git. They will open in whatever replaces Obsidian in fifteen years. No vendor’s export queue is on the critical path of your professional memory.

FAQ

Does Obsidian have AI meeting notes built in?

No. Obsidian ships no AI features of its own and no audio capture. AI meeting notes in Obsidian always come from something outside it: a community plugin, a desktop app that writes into your vault, or a cloud service synced down. The three shapes are compared above.

Does Fazit require an Obsidian plugin?

No. Fazit writes standard Markdown files into a folder in your vault. Obsidian indexes them natively; Dataview and search work immediately.

Where exactly do notes go?

Into a Fazit/ subfolder of the vault you choose in Settings, named yyyy-MM-dd HHmm Title.md.

If I sync my vault, isn’t the transcript in the cloud anyway?

That’s your choice to make — and that’s the point. The text syncs under your sync tool’s encryption and your account. The audio was never a file anywhere, so there is nothing riskier than your own written notes in the loop.

Can I use it without Obsidian?

Yes. The notes are plain Markdown; if Obsidian isn’t installed, Fazit opens them in your default Markdown editor.