Alfred

Alfred is a personal assistant desktop app (Tauri 2 + React/TypeScript) that records your meetings and voice notes, transcribes them locally with Whisper, and turns them into structured notes, summaries, and tasks in a plain Markdown vault — with Claude as your AI assistant on top of it all.

Windows and macOS installers (.msi/.exe, .dmg) are published on the Releases page.

Privacy: what's local, and what uses AI

Alfred has two clearly separate parts. Please read this before recording anything sensitive.

🔒 Transcription — 100% local, nothing leaves your machine

Recording and transcription run entirely on-device via whisper.cpp. Your microphone/system audio and the raw transcript it produces are never sent anywhere for this step — there is no network call involved in turning speech into text.

🤖 AI features — send text to Claude (Anthropic)

Everything beyond the raw transcript is AI-assisted, meaning the relevant text (never raw audio) is sent to Claude to be processed. That covers:

1. Meeting summaries — the transcript is sent to generate the structured write-up (alfred-intelligence/).
2. Task extraction — action items pulled out of the same transcript.
3. Home page daily brief ("today's tasks") — your recent notes/tasks are sent to generate the summary shown on the home screen.
4. Chat with Alfred — your question plus the relevant note excerpts it finds (RAG) are sent to answer with sources.
5. Chat agent actions — from that same chat, Claude can also create, edit, or archive notes and tasks on your behalf (never a hard delete).
6. Onboarding voice context — the first recording where you describe yourself/your company is sent to Claude to structure it into Contexte Alfred.md.
7. Glossary generation — the content of Contexte Alfred.md is sent to Claude to derive a list of proper nouns, fed back into Whisper to improve transcription accuracy on names/jargon.

Two ways to access these AI features, both using Claude under the hood:

• Bring your own Anthropic API key — requests go directly from your machine to api.anthropic.com using your own key. Nothing passes through do·now's servers.
• AlfredIA subscription — a managed option with no key to set up. Requests are routed through our proxy (which holds the shared API key) to Claude, and are not used for anything beyond serving that request.

What it does

• 🎙️ Recording — microphone and system audio (Windows: WASAPI loopback, mic + system mixed by default), or import an existing .wav file.
• 📝 Local transcription — on-device with Whisper, see above.
• 🤖 AI notes & tasks — Claude turns a transcription into a structured summary and a task list, with a review screen to correct anything before it's finalized.
• 🗂️ Markdown vault — everything lives as plain .md files you own (alfred-raw/ for raw transcriptions, alfred-intelligence/ for AI-generated notes and Todo.md), organized by project, with wikilinks and a notes graph.
• 💬 Chat over your notes — ask Alfred questions about your own notes (RAG), get answers with clickable sources.
• 🔗 Sharing — share a note or your task list as a public read-only link.
• 🌐 FR / EN — full UI and AI-output localization.
• 🔑 Bring your own key or subscribe — use your own Anthropic API key, or subscribe to AlfredIA (managed proxy, no key to manage).

Building from source

Windows

Whisper (local transcription, CPU) is on by default — the app builds whisper.cpp via the whisper Cargo feature, part of default.

Prerequisites

• Rust (MSVC toolchain) — install via rustup. Accept the default x86_64-pc-windows-msvc toolchain.
• Visual Studio Build Tools with the "Desktop development with C++" workload (provides the MSVC linker Rust needs).
• WebView2 runtime — preinstalled on Windows 11. If missing, install the Evergreen runtime from Microsoft.
• Node.js 18+.
• CMake — winget install Kitware.CMake (builds whisper.cpp).
• libclang (for bindgen) — easiest without admin: pip install --user libclang. Alternatively install LLVM (winget install LLVM.LLVM, requires admin).

Don't want to install CMake/libclang right now? Skip Whisper for a session with ./scripts/dev-windows.ps1 -NoWhisper (recording still works, no transcription).

Setup — from the repo root, in PowerShell:

./scripts/setup-windows.ps1   # installs sqlx-cli, creates the compile-time dev DB + migrations
npm install
./scripts/dev-windows.ps1     # auto-discovers libclang.dll, runs `tauri dev`

setup-windows.ps1 creates src-tauri/alfred_dev.db and runs the migrations so the compile-time sqlx::query! macros can verify against the schema. That DB is used only at build time — at runtime the app creates its own database in the Windows app-data directory (%APPDATA%\com.alfred.app).

The first build is slow (it compiles whisper.cpp). The CPU backend works on any machine. To use your GPU instead (CUDA/Vulkan), enable the matching whisper-rs feature in src-tauri/Cargo.toml and install that SDK — not covered here.

Model — the small model is bundled into release builds — it works offline from the first launch, no download needed. In dev, if it's not already fetched, download it from Settings → Transcription (or run ./scripts/fetch-whisper-model.ps1 to fetch it into src-tauri/models/ ahead of time). Models are cached in %APPDATA%\com.alfred.app\models\.

Packaging (release build)

./scripts/build-windows.ps1

Fetches the small model into src-tauri/models/ (skipped if already present — that path is gitignored, so every machine fetches it once), then runs tauri build. Installers land in src-tauri\target\release\bundle\ (msi\ and/or nsis\). Release builds published on GitHub are Authenticode-signed; local builds are not unless you have your own signing secrets.

macOS

Whisper is on by default here too (needs cmake + Xcode CLT, both already required for macOS builds).

npm install
npm run tauri:dev                 # dev, with Whisper
./scripts/fetch-whisper-model.sh  # optional: pre-fetch the `small` model into src-tauri/models/
./build-macos.sh                  # release build (.app + .dmg), fetches + bundles `small`, then builds

Audio system capture

• Windows: system_only / mixed recording sources are implemented via WASAPI loopback — no extra setup needed.
• macOS: not implemented yet (ScreenCaptureKit helper, tracked separately). system_only/mixed return an error; use mic_only.

Known Windows follow-ups

• Ingest ("run Claude" feature): resolves the claude CLI from PATH; a .cmd shim may not resolve. Only affects the manual ingest action.

Project structure & specs

The functional and technical source of truth lives in spec/ (one file per module, with a status index). The v1 backlog is tracked in ROADMAP.md. The AlfredIA backend (Claude proxy + metrics, optional subscription) lives in a separate private repository and isn't part of this codebase.

Contributing

Contributions are welcome — see CONTRIBUTING.md for how to propose changes, the branch/PR workflow, and coding conventions.

License

Licensed under the Apache License, Version 2.0.

---

Alfred is built by do·now.