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:

This is the beginning of the README. Read the rest at the original. Read the full README ↗