Steno

Status: release candidate. v1 is code-complete and shipping as pre-releases while the last manual checks on real hardware run. Track the first stable release in issue #74. The full scope and every settled decision live in .plans/2026-09-24-initial-scope.md.

Why Steno

Most meeting recorders join your call as a participant, upload the audio to their servers and keep your notes behind their login. Steno takes the opposite route:

	Hosted bot recorders	Steno
Captures audio by	A bot joins the meeting	Core Audio taps the apps already playing on your Mac
Everyone on the call sees	"Notetaker has joined"	Nothing
Audio is processed	In the vendor's cloud	On your Mac, after the meeting
Notes live in	The vendor's app	Plain files in your Obsidian vault
Works for	Supported video platforms	Zoom, Teams, Meet, FaceTime, Slack, phone calls via Continuity, in-person meetings
Cost	Monthly subscription	Free, open source, your own LLM endpoint

It was built to match the day-to-day quality of commercial meeting recorders for one person and their friends, with files instead of lock-in.

Install

Steno is a signed and notarized Steno.app for macOS 15 or later on Apple Silicon. Pick one of the three routes; all three install the same release build.

Homebrew (recommended, keeps itself updated through Sparkle)

brew tap nicolaischmid/tap && brew install --cask steno

Direct download

Grab Steno-<version>.dmg from the latest release and drag the app to /Applications.

Nix (aarch64-darwin, installs the released app unchanged)

nix profile install github:NicolaiSchmid/steno#steno

On first launch Steno walks you through the three permissions it needs: system audio capture, microphone, and optionally calendar access for meeting titles and attendee names. Then point it at your Obsidian vault and an OpenAI-compatible endpoint, and you are done.

The tap (NicolaiSchmid/homebrew-tap) follows every release, release candidates included. The installed app updates itself through Sparkle, which offers stable releases only. With nix-darwin's Homebrew module:

homebrew.taps = [ "nicolaischmid/tap" ];
homebrew.casks = [ "nicolaischmid/tap/steno" ];

flake.nix installs the released Steno.app unchanged (no rebuild, no re-signing) on aarch64-darwin.

With nix-darwin, add the flake as an input and put steno.packages.aarch64-darwin.steno in environment.systemPackages; nix-darwin links Applications/Steno.app into /Applications/Nix Apps. With home-manager use home.packages and link the app yourself (home.file into ~/Applications, or the mac-app-util module); home-manager does not link Applications/ on its own. nix run does not apply, the output is an app bundle with no bin/.

The app runs from the read-only Nix store, so Sparkle's "Check for Updates…" can download a new version but cannot install it. Update by bumping release.version and release.hash at the top of flake.nix (the release workflow prints both in its job summary) and rebuilding. To silence the scheduled daily check:

defaults write uno.schmid.steno.mac SUEnableAutomaticChecks -bool NO

nix-darwin users who want in-app updates can use the Homebrew cask through homebrew.casks instead; it installs into /Applications, where Sparkle can write.

How it works

flowchart LR
    subgraph capture["1 · Capture"]
        apps["Zoom · Teams · Meet<br>FaceTime · Continuity call"] -->|Core Audio process tap| them["“them” lane"]
        mic["Your microphone"] -->|echo-cancelled| me["“me” lane"]
        phone["iPhone recorder"] -.->|local network handover| room["in-person recording"]
    end
    subgraph process["2 · Process, on your Mac"]
        stt["Transcribe<br>Parakeet or Whisper"] --> dia["Diarize and match<br>known voices"] --> llm["LLM cleanup, summary,<br>tasks · text only"]
    end
    subgraph deliver["3 · Deliver"]
        vault["Obsidian vault<br>one folder per meeting"]
    end
    capture --> process --> deliver

1. Capture. When another app opens your microphone, a small prompt asks whether to record. Say yes, or start and stop from the menu bar at any time. The other side of the call comes from a Core Audio process tap, your side from the mic, kept as two separate lanes so Steno always knows who is "me" and who is "them".
2. Process. After the meeting, the recording is transcribed on-device with Parakeet TDT v3 or WhisperKit large-v3-turbo, diarized, and matched against the voices Steno has heard before. Only the resulting transcript text goes to the LLM endpoint you configured, which fixes the Denglish, writes the summary from a template and extracts tasks with assignee, priority and due date.
3. Deliver. The finished meeting is pushed into your vault as Markdown, VTT and JSON. Re-export at any time; Steno overwrites only the files it wrote.

What lands in your vault

One dated folder per meeting, named after the title. The folder note is what Obsidian opens when you click the folder.

Meetings/
└── 2026-09-24-produktstrategie-90-10-roadmap-fuer-q4/
    ├── 2026-09-24-produktstrategie-90-10-roadmap-fuer-q4.md              # summary, decisions, scratchpad
    ├── 2026-09-24-produktstrategie-90-10-roadmap-fuer-q4 - Transcript.md
    ├── 2026-09-24-produktstrategie-90-10-roadmap-fuer-q4 - Tasks.md      # Obsidian Tasks syntax
    ├── transcript.vtt
    ├── meeting.json
    └── audio.m4a                                                         # opt-in per meeting

The folder note carries queryable frontmatter and links to the people who were there:

---
title: "Produktstrategie: \"90/10\" & Roadmap für Q4"
date: 2026-09-24T14:00:00
duration: 90
participants:
  - "[[Anna Müller]]"
  - "Jérôme Dupont"
  - "[[Nicolai Schmid]]"
tags:
  - "meeting"
  - "Kunde-ACME"
source: "mac-call"
template: "default"
language: "de"
---

# Produktstrategie: "90/10" & Roadmap für Q4

2026-09-24 14:00–15:30 · 1 h 30 min · Mac call · [[… - Transcript|Transcript]] · [[… - Tasks|Tasks]]

## Summary

### Executive Summary

- **Fokus**: **Anna Müller** setzt 90 Prozent auf den Kern und 10 auf Experimente.

## Decisions

- 90/10-Aufteilung wird umgesetzt.

Tasks come out ready for the Obsidian Tasks plugin, so they show up in your existing queries next to everything else:

- [ ] Angebot an ACME schicken [[Anna Müller]] ⏫ 📅 2026-10-01
- [ ] Roadmap-Folien aktualisieren [[Nicolai Schmid]] 📅 2026-10-15

Summaries follow one of four built-in templates: Default, Customer Discovery, Daily Standup and Interview. Pick a different one after the fact and re-run the summary without touching the transcript.

Features

Records without a bot Core Audio process taps for the other side, your mic for you. Nobody in the call sees a notetaker, and it works for anything that plays audio on your Mac, including phone calls taken via Continuity and in-person meetings.

Everything on your Mac Transcription, diarization and speaker matching run on-device. The only network traffic Steno generates is transcript text to the LLM endpoint you chose.

Remembers voices Name a speaker once and Steno recognises them next time. Unknown speakers get a ten-second clip to listen to, plus name suggestions from the conversation and the calendar invite.

German, English and Denglish Language is detected per meeting, and an LLM cleanup pass fixes the code-switching, casing and product names every speech model gets wrong.

Knows when a meeting starts A prompt appears when another app opens the microphone. Calendar titles and attendees are filled in from EventKit, no Google Calendar API involved.

Files, not lock-in Markdown, VTT and JSON per meeting. Optional per-person pages linking their meetings. Read, search and automate in your vault with whatever tools and agents you already use.

You decide what to keep Audio retention is a setting: delete right after processing, keep for N days, or keep forever, with a per-meeting override.

Bring your own model Any OpenAI-compatible endpoint works for post-processing, from a hosted API to a model running on the same Mac. Base URL and API key, nothing else.

Privacy

Data	Where it goes
Meeting audio	Stays on your Mac, in a folder you choose. Deleted per your retention setting.
Transcript text	Sent to the LLM endpoint you configure, and only there.
Speaker voice embeddings	Stored in the local SQLite database. Never uploaded.
Summary, tasks, notes	Written to your vault. The app never runs git or syncs anything itself.
Phone recordings	Travel from your iPhone to your Mac over the local network only, TLS with pinned keys.

There is no Steno server. Disk encryption is FileVault's job.

Phone companion

For meetings away from the Mac there is a deliberately dumb iOS recorder in mobile/. It records, keeps a local queue, and hands the files to your paired Mac over the local network when both are on the same Wi-Fi. Pairing is a one-time QR scan; transport is HTTPS with pinned certificates. Processing happens on the Mac with the same pipeline as any in-person recording. The phone holds no notes and needs no account.

What Steno does not do

Steno is scoped on purpose. It does not run on Windows, Linux or Intel Macs. There is no live transcript during the call, no chat over your meetings, no editing of generated text inside the app, no team features, no hosted service and no App Store build. The app is the source of truth for your data; reading, searching and automation happen in your vault.

For developers

Swift 6 and SwiftUI on macOS 15+, SQLite via GRDB, FluidAudio and WhisperKit for speech, SpeexDSP for echo cancellation, SwiftNIO for the phone handover, Sparkle for updates. The iOS recorder is an Expo dev-client app.

AGENTS.md           conventions for coding agents (CLAUDE.md links here)
Package.swift       Swift package: StenoCore, StenoAudio, StenoSpeech, StenoLLM,
                    StenoAdapters, StenoHandover and the `steno` CLI
apps/macos/         SwiftUI app over the package, xcodegen project, own README
apps/site/          landing page (steno.nicolaischmid.com), Next.js static export in the root pnpm workspace
mobile/             Expo iOS recorder, own pnpm project, own README
.plans/             scope and dated implementation plans
docs/research/      teardown of commercial recorders and the open-source landscape
.github/workflows/  repository, Swift, web, site and mobile CI; macOS release on `v*` tags

swift build && swift test                 # the package; Apple-only modules need macOS
swift run steno record --help             # record from the CLI
swift run steno dev models list           # speech and diarization model assets
swift run steno dev bakeoff --help        # compare speech engines over a folder

Building the app: apps/macos/README.md. The phone side: mobile/README.md. Conventions, review priorities and the plan workflow: AGENTS.md. Everything that is not UI lives in the package, so the CLI, the macOS app and any future surface share one implementation.

Background

Steno started as an open-source alternative to commercial meeting recorders. Along the way it borrows ideas from anarlog, Parrot and AudioCap, all rebuilt rather than forked. The name is a nod to stenographers and is known to collide with a few commercial products; that is acceptable for a personal open-source project.

License

MIT, see LICENSE, except where a directory or file carries its own licence: crates/steno-audio/vendor/speexdsp (BSD), crates/steno-audio/src/codec/salvage/tables.rs (MPL-2.0, AAC tables copied from the symphonia-codec-aac crate; the AUR package ships a copy, packaging/aur/symphonia-aac-tables.rs).