Local-first dictation, voice notes and meeting memory for macOS. Everything runs on your Mac — no cloud APIs, no subscriptions.

LocalFlow is an open, privacy-first alternative to Wispr Flow built for Apple Silicon. Press ⌘B, speak anywhere, and get clean, edited text pasted into the field you were typing in. Record calls with one shortcut, keep a voice diary, and ask questions across your own searchable archive — with every model running locally.

• Dictate anywhere (⌘B). A floating window shows your words live while you speak — stable text plus a dimmed draft tail, a word counter, and a voice-level waveform that dances with your loudness. Resize it like any window (it remembers position and dimensions). Press ⌘B again and the window stays visible while the transcript is finalized and edited on device, then inserts the result at the cursor and dismisses itself. Escape cancels; the clipboard is restored after pasting. If the target field truly cannot receive text, a small fallback panel keeps the result one click away.
• Shortcuts you actually like. Both global shortcuts (dictation and meeting capture) are re-recordable in Settings with live validation: conflicts with each other and system-critical combinations (⌘C/⌘V/⌘X/⌘Q/⌘W/⌘Space/⌘Tab) are rejected, function keys work without modifiers, and ⌘-shortcuts trigger on the left ⌘ only — the right ⌘ keeps reaching applications, so ⌘B still means Bold.
• Two editing strengths. Tidy removes speech noise while preserving facts, numbers, negations and your tone. Flowing text reorganizes a stream of thoughts into coherent paragraphs without inventing anything. A free-form your style field tunes both.
• Voice notes as a quiet diary page. Create a note, write a title and a line of context, then dictate into it whenever you're ready. Notes can be extended by voice later.
• One-press meeting capture (⇧⌘M). Records system audio and your microphone as separate tracks, transcribes both into speaker turns, splits speakers post-hoc, and condenses the transcript in two stages — notes per section, then one merged summary with decisions, open questions and [mm:ss] timestamps that play the matching audio. Measured on a real 29-minute two-person call: 21 892 characters of speech to a 6 095-character summary carrying 44 verified timestamps (docs/validation.md). Two honest limits: the system track captures every app's audio, so a video played during a call becomes a phantom participant, and diarization can still split one voice into two identities.
• A memory that is actually yours. A personal dictionary (heard → preferred spelling) is applied before every edit; familiar voices can be saved and recognized across meetings. Ask the archive a question and every claim comes with a citation and a playable timestamp.
• Engineered for low load. Models load on demand and unload automatically after idle (ASR after 2 min, editor after 1 min by default). A dictated sentence edits in ~1–2 s on an M3 Pro; nothing runs in the background while you're silent.

All weights are pinned to exact revisions with per-file checksums and downloaded once from their publishers; after that LocalFlow works fully offline.

The editor was chosen by a measured bake-off on a 10-case Russian dictation corpus (docs/benchmarks): Gemma 4 E2B is ~2× faster than Qwen3-4B (0.32 s first token, ~76 tok/s) at equal or better fact preservation, and correctly handles self-corrections like "send 12 files, not 15". Qwen3.5-4B was disqualified for silently changing meaning. Both remaining editors are selectable in Settings.

Safety rails. Every edited draft passes deterministic guards (number multisets, negation counts, corrected-number patterns, length ratios). A suspicious edit is never silently trusted: the minimal deterministic cleanup is delivered instead, and the model's proposal is kept as a separate version. The original transcript is always preserved. Archive answers without valid citations are rejected.

Requirements: Apple Silicon Mac, Xcode 26+ with the Metal toolchain, XcodeGen.

brew install xcodegen
xcodebuild -downloadComponent MetalToolchain
Scripts/build.sh     # xcodegen + xcodebuild Release + local signing
Scripts/test.sh      # 55 unit/interface tests
Scripts/install.sh   # installs to ~/Applications/LocalFlow.app

First run: open Settings → Models, download the recognizer and editor, grant Microphone and Accessibility (for ⌘B), and — for meeting audio — Screen & System Audio Recording. The app UI is currently in Russian; the codebase and docs are English-friendly and localization is on the roadmap.

The build is signed with a persistent local certificate (stored outside the repo), so permissions survive re-installs on your own Mac. This is not a notarized distribution — see docs for the honest limitations list.

Everything lives in ~/Library/Application Support/LocalFlow/: a SQLite archive with full-text search (archive.sqlite), chunked 30-second audio parts with a crash-safe journal, and checksum-verified model weights. Audio is pruned after 30 days (pinned items are kept), texts stay until you delete them, and there is no telemetry of any kind. Recording survives crashes: transcription is checkpointed per window and resumes where it stopped.

Sources/LocalFlowCore/   engine: audio capture, ASR/LLM/diarization, safety guards, SQLite store
Sources/LocalFlow/       SwiftUI app: overlay panel, archive, diary, settings
Sources/LocalFlowBench/  CLI benchmarking harness (ASR/editor/diarization pipelines)
Scripts/                 build, test, install, signing, model pinning, editor bake-off
docs/                    architecture, plan, validation log, benchmarks, product strategy

v0.2.0 — daily-driver quality for the author's setup (M3 Pro / 18 GB, macOS 26): dictation, notes, meetings, archive Q&A, model choice. Known limits: diarization can merge similar voices; latency targets on real speech are still being measured; calendar/Slack/MCP integrations and mobile are future work. Full details: docs/plan.md, docs/validation.md, docs/product-strategy.md.

Code: MIT. Model weights keep their publishers' licenses — see THIRD_PARTY_NOTICES.md.