cereal

Cereal

First Wispr Flow, then Obsidian. Now we're making Granola free. Stop paying for dumb shit.
The free, local AI notepad for your meetings. No bot joins your call.
Jot a few words during the meeting. When it ends, the AI on your computer turns them and the transcript
into clean notes with decisions and action items. No account. No limits. Nothing is sent anywhere unless you turn on sync,
a cloud AI option or a connector.

_{Inspired by Granola
Cereal is an independent open-source project and is not affiliated with or endorsed by Granola.}

Cereal (formerly Toolbay Notes) is part of Toolbay: free, open-source rebuilds of paid software. It records your microphone (you) and your computer's sound (everyone else) as two tracks, so it works with Zoom, Google Meet, Microsoft Teams, Slack huddles, Discord, Webex or a phone app without inviting anything into the call. Speech recognition (whisper.cpp), speaker separation (sherpa-onnx, pyannote segmentation) and the notes AI (llama.cpp with an open Qwen model) all run on your own hardware.

Written from scratch (Tauri 2 + Rust + React). No code, assets or services from any commercial product are used.

Install

Windows 10/11: download Cereal_x.y.z_x64-setup.exe from Releases and double-click it. No admin prompt, nothing else to install: the speech, speaker and notes engines and a starter speech model are inside. Cereal picks the best AI models for your computer and fetches them in the background.

macOS 13+: download Cereal_x.y.z_aarch64.pkg (Apple silicon) or _x64.pkg (Intel) and open it. Cereal isn't notarized by Apple yet, so macOS may say it's from an unidentified developer: System Settings › Privacy & Security › Open Anyway. Allow Microphone and "Screen & System Audio Recording" when asked (only the sound is used, never your screen). Not yet tested on a real Mac. (The .dmg works too, but macOS calls a downloaded un-notarized app "damaged"; xattr -dr com.apple.quarantine /Applications/Cereal.app fixes that.)

Through the Toolbay app: find Cereal and click Install. One step, models included.

How it works

1. A call starts. Cereal notices (Zoom, Meet, Teams, Slack, Discord, Webex, browser calls…) and starts taking notes, with a small floating indicator (Pause, Stop). Prefer to be asked? Settings › When a meeting starts. With a calendar connected, you also get a nudge a minute before each meeting.
2. The first time, it reminds you to tell the others (in many places everyone must agree to be recorded) and gives you a message to paste in the chat.
3. Type rough notes as the call goes. The live transcript fills in on the right.
4. When the call ends, Cereal stops by itself, re-transcribes at full quality, tells the other speakers apart, removes echo, and writes your notes with your template. Your own points stay dark; what the AI added is grey. Click the magnifier next to any line to see where it came from in the transcript.
5. Chat with one meeting, a folder, a person, a company, selected notes or everything (Ctrl+J), with sources. Search everything (Ctrl+K). Share as Markdown, Word, PDF or email, or auto-post a folder to Slack or any webhook.

Short calls with no talking are thrown away automatically.

Speed (RTX 4070 Ti, 44-minute 4-person meeting): transcript and speakers ready ~60 s after the call, notes ~22 s later. CPU only (same PC): transcript written live during the call, notes ~9 minutes after.

Features

• Two-track capture, no bot; live transcript; full-quality pass after the call; speaker separation; echo removal; auto-gain for quiet calls; follows device changes; never mutes other audio.
• Auto-detect and auto-record meetings (or ask, or do nothing), per-app rules and "ignore this app", auto-stop, false-alarm cleanup, always-visible indicator.
• Calendar via your private iCal link (Google, Outlook, iCloud): Coming up, pre-meeting nudge, prep briefs from past notes with the same people, notes titled with the event and its attendees.
• Granola-style notepad: Markdown, checkboxes, images (paste or drop), @mentions, / for templates.
• 26 templates (1:1, Sales call, Discovery, Stand-up, Interview, Retro, Board meeting…) plus your own, any as default.
• Chat with scopes and numbered citations, / recipes (12 built-in plus your own), voice questions, file attachments.
• Spaces: My notes and folders with icons, descriptions, files as chat context, auto-filing and webhooks; People and Companies views; Trash with 30-day restore.
• Import audio files (MP3, M4A, WAV, FLAC…) as notes. Export one note (Markdown, Word, PDF) or everything (CSV, JSON).
• Vocabulary, profile and notes language; transcript retention; light and dark; Toolbay account sync; in-app feedback.
• Your notes in other AI apps: a local MCP server (cereal --mcp, for Claude Desktop, Cursor…) and a CLI (cereal notes list | get | transcript | search). Send to Notion, log in HubSpot, post to Slack, with your own keys.
• Floating indicator with screenshot and live captions, Alt+Space quick chat, mini window, chimes, cereal:// links.
• Automatic model choice; optional cloud engines with your own key. No telemetry.

The full checklist against Granola is in docs/FEATURES.md.

Privacy

Audio, transcripts, notes and the search index live in your app data folder (%APPDATA%\ai.toolbay.cereal on Windows, ~/Library/Application Support/ai.toolbay.cereal on Mac). Network use: downloading AI models, your calendar feed, Toolbay sync only if you turn it on (off by default), feedback you choose to send, connectors and webhooks you set, and a cloud engine only if you add your own key.

Sync sends your profile (name, job title, role, about you, company, vocabulary), consent note, recording and notes choices and custom templates to your Toolbay account on toolbay.ai. Toolbay's server can read this data; it is not end-to-end encrypted. Calendar (iCal) links, recordings, transcripts, notes and API keys are never synced. Turn sync off or delete the synced copy any time in Settings > Account.

Build from source

npm install
npm run tauri dev

npm run tauri build makes the installer. Tests: cargo test --workspace. See AGENTS.md.

License

MIT. whisper.cpp, llama.cpp and sherpa-onnx are MIT/Apache-2.0; Whisper models (MIT), Qwen3 (Apache-2.0), pyannote segmentation 3.0 (MIT), 3D-Speaker CAM++ (Apache-2.0). Fonts: Newsreader and Inter (OFL).