How to Turn Meeting Transcripts into an Obsidian Second Brain (Fully On-Device)

Published September 05, 2026

September 5, 2026 · 10 min read

If you already keep an Obsidian vault, the fastest way to make meetings searchable is to route every transcript into it. The question in 2026 is not whether to do that — it is how to do it without every conversation making a detour through a vendor's cloud first. This guide walks through a fully on-device workflow: capture the meeting locally on your Mac or iPhone, export a Markdown transcript, and drop it into an inbox folder inside your Obsidian vault. No cloud transcription API, no meeting bot, no vendor-side copy.

The workflow has three moving parts: a capture app that transcribes on-device, an inbox folder inside your vault, and a routing step (a shell rename, an Apple Shortcut, or an Obsidian plugin) that gets the file where it belongs. Everything downstream — templates, backlinks, Dataview queries, PARA sorting — is standard Obsidian, and it works whether your capture tool is Basil AI, a Whisper wrapper, or Apple's built-in Voice Memos transcription.

Why route meeting transcripts through Obsidian at all?

The value of a personal knowledge base is compounding retrieval: a decision you made in a January product review becomes searchable evidence in July's postmortem. Meetings are where most of that decision-making happens, but they evaporate the moment the call ends unless somebody captured them. Obsidian is a natural destination because it stores plain Markdown files locally, links between them, and never uploads your vault unless you explicitly configure a sync provider.

The rub is that the dominant meeting-capture tools were built cloud-first. Shadow's review of Obsidian meeting workflows notes that Granola's documented Obsidian integration relies on either a manual "Copy text" action that puts the AI-enhanced summary on the clipboard as Markdown, or a paid Zapier trigger that exposes summary and transcript. Both approaches assume the transcript already exists on a vendor server. That is fine for meetings with low sensitivity, and inconvenient for anything you would rather never leave your machine.

The on-device workflow in one page

Here is the entire pipeline, top to bottom, before we drill into each step:

1. Capture — Record and transcribe the meeting locally on your Mac or iPhone using an on-device app.
2. Export — Write the transcript as a dated Markdown file (e.g., 2026-09-05 - Product Review.md).
3. Route — Send the file to a single inbox folder inside your Obsidian vault (e.g., 0 - Inbox/).
4. Enrich — Add frontmatter (attendees, project, tags) either at export time or via an Obsidian template.
5. Sort — Move the note to its permanent home (Project, Area, Resource) using PARA, Johnny Decimal, or Zettelkasten conventions.

Steps 1–3 are automatable. Step 4 usually is too. Step 5 is deliberately manual for most people, because it is where you decide what a meeting meant.

Step 1: On-device capture (Mac and iPhone)

Apple ships an on-device speech recognizer on every current iPhone, iPad, and Apple Silicon Mac. Apple's Speech framework exposes it to any app that wants to use it, and starting with iOS 26 the newer SpeechAnalyzer API produces higher-quality transcripts entirely on device. Third-party apps built on those APIs — Basil AI on Mac and iPhone, Aiko, MacWhisper, Yaps — never need to upload audio to a cloud endpoint. For a walkthrough of the newer API, see our deep dive on iOS 26 SpeechAnalyzer and on-device transcription.

You can also roll your own with local Whisper models. A homegrown workflow documented at heliomass.com combines a local Whisper transcription pass with a summarization step running through Ollama (Qwen3 in that setup), all on-laptop, and writes the results into an Obsidian vault with a date-prefixed filename for chronological ordering. The author's reasoning is worth quoting because it is exactly the reasoning that drives most Obsidian users to this workflow: they want sensitive data to remain in a secure enclave, and their laptop is that enclave.

What to record from

The choice of capture source matters more than the choice of transcription engine:

• In-person meetings — Record with your device microphone. Basil AI supports up to 8 hours of continuous recording, which is long enough for a full-day workshop.
• Remote calls (Zoom, Teams, Meet) — Capture system audio on your Mac. This is the same architecture the Granola lawsuit centers on, so if the meeting is external, tell participants you are recording before you start. On-device processing does not by itself resolve consent obligations.
• Teams meetings you host — Teams can export a WebVTT transcript directly, which skips the audio-to-text step entirely. Point your workflow at both audio files and .vtt files so it handles either input.

Step 2: Export a clean Markdown file

The transcript file that lands in Obsidian should be Markdown, dated, and named in a way that sorts chronologically. A reliable pattern is YYYY-MM-DD - Meeting Title.md. If your capture app produces raw transcript only, wrap it with a summary section at the top using a local LLM (Ollama running Llama, Qwen, or Mistral works well on M-series Macs).

The Obsidian community has converged on a canonical structure for meeting notes:

---
date: 2026-09-05
attendees: [Alex, Sam, Priya]
project: "[[Q4 Launch]]"
tags: [meeting, product-review]
---

## Summary
- Decision to ship on Oct 14.
- Alex owns the pricing page rewrite.

## Action items
- [ ] Alex: pricing page draft by Sep 12
- [ ] Priya: legal review by Sep 20

## Transcript
(full text)

The frontmatter block is what unlocks Obsidian's structured queries. Obsidian Properties, introduced in version 1.4 in August 2023, gave frontmatter a first-class UI, and the Dataview community plugin turns those properties into a query language. Together they mean that once your transcripts land with clean frontmatter, filtering "every meeting with Alex about Q4 Launch" becomes a one-line query.

Step 3: Route the file into an inbox folder

On macOS this is trivial: point your transcription app's export path at a folder inside your Obsidian vault, for example ~/Obsidian/MainVault/0 - Inbox/. Every completed transcript appears in that folder when the meeting ends. As Shadow's PARA-method guide puts it, the pipeline lands the file automatically — no copy-paste, no export button, no "share with Obsidian" plugin to maintain.

On iOS the routing is a little more work because Obsidian does not natively support Apple Shortcuts. Two solutions:

1. iCloud Drive vault — Store your Obsidian vault inside iCloud Drive and have your transcription app write directly to Obsidian/MainVault/0 - Inbox/. Files sync to your Mac automatically.
2. Apple Shortcuts + a bridge app — Use Actions For Obsidian, a third-party app that exposes over 40 Shortcuts actions for creating, appending, and organizing notes in your vault. Build a share-sheet shortcut that receives the Markdown export from your transcription app and calls the "Append Text to a Note" action.

The n8n community workflow for on-device iOS transcription follows the same pattern: record a Voice Memo, run the on-device transcription Shortcut, then post the Markdown through a webhook that writes it into a folder synced with your Obsidian vault. The author flags a real limitation — transcriptions longer than about an hour through the Voice Memos Shortcut can produce unstable results — which is one reason purpose-built streaming apps outperform after-the-fact batch transcription for anything past a short call.

Step 4: Enrich with templates and templater-style automation

Once the raw transcript is in Obsidian, a template can automatically:

• Prepend the frontmatter block with today's date and default tags.
• Insert a ## Summary, ## Decisions, and ## Action items scaffold.
• Wikilink attendees to their existing people pages.
• Backlink the note to the relevant Project MOC (map of content).

The Meeting Notes Obsidian plugin automates a lot of this at capture time — it can generate diarized transcripts, summary-only notes, or full meeting notes with default Markdown subheaders per selected section, and format task-style items as - [ ] Follow up with Alex checkboxes. If you use that plugin, be aware its default summarization sends the transcript to OpenAI; you can point it at a local model via Ollama-compatible endpoints if you want the whole pipeline to stay on device.

Step 5: Sort with PARA, Johnny Decimal, or Zettelkasten

Once a transcript is in 0 - Inbox/, the sorting decision is: does this belong to an active Project, an ongoing Area of responsibility, a Resource you might reference later, or an Archive of finished work? That is Tiago Forte's PARA method, and it works especially well for meeting notes because most meetings map cleanly to a single project.

Johnny Decimal and Zettelkasten are alternatives — the first uses numeric folder IDs, the second uses atomic, densely-linked notes instead of hierarchy. Both work with the same capture-and-route logic; the AI capture step does not care which organizational system you use, as long as the inbox folder exists.

On-device vs cloud handoffs: a side-by-side

Every AI notetaker that markets an Obsidian integration falls somewhere on the spectrum below. The relevant question for a second brain is where the transcript exists before it lands in your vault.

Dimension	Cloud AI notetaker → Obsidian	On-device app → Obsidian
Where audio is processed	Vendor server (US, EU, or Frankfurt)	Your Mac or iPhone; no upload
Where transcript first exists	Vendor database	Local filesystem inside your vault
Handoff mechanism	Copy-paste, Zapier, webhook, or REST plugin	Direct file write to vault folder
Vendor retention window	Days to indefinite (varies by plan)	None — no vendor copy exists
Training-data use	Often opt-out by default	N/A — vendor never sees the data
Subpoena target	Vendor can be compelled to produce	Only you hold the file
Offline capability	Requires internet at capture time	Works fully offline
Data ownership	Governed by vendor ToS/DPA	Same as any other file in your vault

The right-hand column is the reason most privacy-minded Obsidian users prefer an on-device pipeline. It is not that cloud tools cannot land Markdown in a vault — SystemSculpt's Obsidian transcription workflow documents a solid managed-service setup — it is that the vendor server exists as a layer between the microphone and the file. For sensitive meetings that layer is a liability.

What the cloud alternatives actually offer

To be fair to the ecosystem: Granola and Jamie both offer real Obsidian handoffs. Granola combines device-audio transcription with the notes you type during a meeting, and its sharing documentation says Copy text puts the AI-enhanced summary on the clipboard as Markdown while the full transcript is copied separately from the transcript panel. Paid desktop users can also use Granola's Zapier integration, whose trigger exposes the summary in Markdown plus the transcript and note link.

Jamie captures meetings without a bot on macOS, Windows, and iOS. Its data-handling documentation says audio is uploaded for processing in Frankfurt and deleted after processing, while the transcript remains stored. That is a legitimate architecture — but note that "uploaded for processing" is exactly what an on-device workflow avoids.

The Chamberlain v. Granola class action filed July 30, 2026 in the Northern District of California is worth reading before you commit to any bot-free cloud transcriber, because the complaint centers on the invisibility of the capture, not just the cloud storage. Our detailed writeup covers the case in our article on the Otter.ai ruling and vendor-side eavesdropper theory.

How Basil AI fits into this workflow

Basil AI is designed as the capture end of exactly this pipeline. On both iPhone and Mac, it uses Apple's on-device machine-learning stack to transcribe audio locally, then exports the transcript as Markdown you can drop into your Obsidian vault the same way you would drop any other file. Because the transcription runs on the Apple Neural Engine, the audio never leaves the device — there is no server-side copy for a vendor to retain, train on, breach, or produce under subpoena.

Concretely, three product decisions make Basil AI a natural fit for an Obsidian workflow:

• 8-hour continuous on-device recording. Long enough for board meetings, all-day offsites, and multi-track workshops without hitting the ~1-hour ceiling that trips up Voice Memos-based Shortcuts.
• Markdown export with dated filenames. Files are ready to write straight into 0 - Inbox/ — no post-processing required to make them sort chronologically.
• Apple Notes + iCloud integration. If you use Apple Notes as a staging layer before Obsidian, Basil AI drops summaries there first and syncs across your devices before you promote the note into the vault.

None of this is a compliance claim. On-device processing is an architecture fact — the vendor does not hold your audio because there is no vendor server in the loop. Whether that architecture satisfies HIPAA, GDPR, the SEC's Advisers Act requirements, or your firm's DPAs is a determination for your compliance officer. On-device just removes one large category of vendor-side risk from the analysis.

A checklist before you commit to a workflow

1. Does the capture app run transcription on-device, or does it upload audio? Read the privacy policy, not the marketing page.
2. Does the app export Markdown directly, or only through a paid integration tier?
3. Can you point the export path at a folder inside your Obsidian vault without a bridge service?
4. If you use iOS, does the app work with Apple Shortcuts or a bridge like Actions For Obsidian?
5. For long meetings, does the app handle multi-hour capture in a single file, or does it split and lose context?
6. If the meeting involves external parties, have you disclosed the recording? On-device processing does not resolve consent obligations under two-party statutes.
7. Does your vault sync method (iCloud, Syncthing, Git) meet your own data-handling requirements? An on-device transcript that syncs to an unencrypted cloud is only half-private.

Common pitfalls

Assuming "local" means "private." Some tools transcribe on your machine but call a cloud LLM for the summary. If the summary step ships your transcript to OpenAI, the transcript is now on OpenAI's servers. Keep the summarization step local too — Ollama with a mid-size model handles this well on Apple Silicon.

Skipping the inbox pattern. Writing every transcript directly to its final Project folder feels efficient until you get a meeting that touches three projects and you cannot decide where it goes. The inbox is the pause that lets you decide with a clear head.

Over-templating. If your template has 12 required fields, you will stop filling them in by week three. Start with three fields — date, attendees, project — and add more only when a query you actually run needs them.

Ignoring the compliance layer. On-device removes the vendor-side risk. It does not remove GDPR Article 5 obligations around lawful basis, minimization, and retention, and it does not change consent requirements under state wiretap laws.

Where this workflow shines vs where it does not

An on-device Obsidian pipeline is the right default when you own your device, your meetings are sensitive, and you already think in Markdown. It is worse than a managed cloud service when you need collaborative live editing during the call, cross-organization sharing with people who do not use Obsidian, or a search index across a whole company's meetings.

For a fuller comparison of the current AI notetaker landscape and where on-device versus cloud tools each win, see our AI meeting notetaker comparison guide. If your primary concern is competitor privacy behavior specifically, the Otter.ai privacy policy and Fireflies privacy policy are both worth reading in full before you route a single meeting through them.

The one-hour version of this project

If you want to try this today, block one hour and do the following in order:

1. Install Basil AI on your Mac and iPhone.
2. Create 0 - Inbox/ inside your existing Obsidian vault.
3. Point Basil AI's export path at that folder (Mac) or store your vault in iCloud Drive so iOS exports sync (iPhone).
4. Record a 5-minute test meeting (talk to yourself about what you want your second brain to do).
5. Confirm the Markdown file lands in 0 - Inbox/ with a date-prefixed filename.
6. Move it to a Project or Area folder. Done.

Everything else — Dataview queries, PARA sorting, wikilinked attendee pages, template automation — is optional polish on a workflow that already works after step 5.