Local speech-to-text menu bar app for macOS. Dictate anywhere, anytime — the transcribed text pastes directly into whatever app has focus: Mail, Slack, LinkedIn, Notes, WhatsApp Web, your terminal, your IDE.
No always-on cloud listener. No server round-trip for the raw transcription. Press a hotkey, speak, press the hotkey again — text appears.
⚠️ Platform support: macOS only (Apple Silicon)blitzbot is not available for Windows or Linux. The whole app is deeply integrated with macOS-specific APIs: SwiftUI + AppKit UI, Carbon global hotkeys, CGEvent Cmd+V simulation, TCC accessibility permissions, macOS Keychain. Porting isn't a recompile — it would be a rewrite of every OS-touching layer.
The reusable pieces are cross-platform: whisper.cpp runs on Linux and Windows, and the Claude API is HTTP. If you want to build a Windows or Linux equivalent using those building blocks, go for it — the prompts in
Sources/blitzbot/Mode.swiftare portable. PRs adding a separate cross-platform implementation (Tauri, Electron, native Win/Linux) are welcome as a sibling directory, but will not be maintained by me personally. See the Contributing section.
Inspiration: Christoph Magnussen's video "Nie wieder Tippen! Meine eigene Speech-to-Text App (Claude Code)" — https://www.youtube.com/watch?v=vVTl1dqPL0k
This is my own take on the idea: different architecture, different modes, different name. The point — as Christoph put it — is the shift from tool tourist to application master.
- Speech-to-text: whisper.cpp running locally. Offline. Private. No audio leaves your machine.
- Text polishing (optional, per mode): Anthropic Claude API. Only the transcribed text is sent, never the audio.
- One API key to set up: or more, via Connection Profiles — each profile holds a provider, base URL, auth scheme, and key. Switch profiles with one click.
- Open source: MIT license.
- The eight modes
- Office Mode — interactive selection-rewriter
- Privacy Mode — local PII anonymization
- Connection recovery — no transcript left behind
- Connection profiles
- Installation
- First launch on macOS (Gatekeeper workaround)
- Usage
- Settings
- Data flow & privacy
- Cost
- macOS permissions explained
- Troubleshooting
- Development
- Architecture
- Contributing
- Roadmap
- Changelog
- License
Modes 1–7 are voice-driven (mic + Whisper + paste). Mode 8 (Office) is the first non-voice mode: a dedicated window that accepts typed text or a dropped file.
| # | Mode | Default hotkey | Tagline | Behavior |
|---|---|---|---|---|
| 1 | Normal | ⌘⌥1 |
Voice in. Text out. | Raw Whisper transcript. No cloud call by default. Zero cost. (Optional: enable STT Correction in Settings → General to run a dialect-fix LLM pass on the transcript before pasting — adds one cloud call.) |
| 2 | Business | ⌘⌥2 |
Voice in. Business-ready out. | Claude rewrites into clear, polite, structured business communication (emails, customer replies, LinkedIn posts). |
| 3 | Plus | ⌘⌥3 |
Speak in writing. | Claude removes filler words (uhm, also, you know) and fixes grammar — your voice stays intact. Not a business makeover. |
| 4 | Rage | ⌘⌥4 |
Frustration in. Calm out. | Claude strips insults and aggressive tone — the substance of your criticism stays sharp. Good for writing angry emails you won't regret. |
| 5 | Emoji | ⌘⌥5 |
Voice in. Text with emojis out. | Original wording 1:1, dotted with tasteful emojis (roughly 1 per 1-2 sentences). |
| 6 | Prompt | ⌘⌥6 |
Idea in. Prompt out. | Dictate a loose idea — Claude turns it into a clean, precise prompt you can paste into any AI tool (ChatGPT, Claude, Claude Code, Cursor, Aider, Copilot, Gemini, …). Output is the prompt itself, not the result. |
| 7 | Translate | ⌘⌥7 |
DE in. EN out. (Or the other way round.) | Auto-flip translator. Whisper detects the source language; Claude translates into the other one (DE→EN or EN→DE), preserving register and tone. Names and code identifiers stay unchanged. |
| 8 | Office | (no default — set in Settings → Hotkeys) | Selection in. Choice + paste out. | Interactive selection-rewriter. Grabs the currently selected text, shows it in a preview, lets you pick any voice-mode prompt (Business, Plus, Rage, Emoji, Prompt, Translate), override profile + model, tweak the text, then pastes the result back into the source app on ⌘↵. File-drop also works as an alternative input. See Office Mode. |
Settings → General → Output language:
- Auto (default) — Whisper detects the language of what you spoke; the Claude polish runs in the same language
- German — forces DE transcription + polish regardless of what you said
- English — forces EN transcription + polish
While recording, the HUD shows a small DE/EN badge next to the mode name so you see what the app will actually output.
Start a recording with ⌘⌥1 (Normal), change your mind halfway through, press ⌘⌥4 while still speaking — the recording keeps going, but will be processed as Rage when you stop. The floating HUD reflects the current mode live.
Or click directly: the HUD has six mode pills at the bottom (Office is excluded from the HUD — it has its own window). Click any pill during recording to switch live. The Stop button on the right ends the recording with a mouse click — useful when your keyboard is full of other input and you don't want to hit the hotkey.
- Hotkeys: Settings → Hotkeys → click the recorder next to each mode, press your desired combo.
- Prompts: Settings → Prompts → edit the Claude system prompt per mode. Make Business more formal, Rage softer, Emoji denser.
- Vocabulary: Settings → Vocabulary → add proper nouns and jargon. They're fed to Whisper as context so it spells Anthropic, Kubernetes, your colleagues' names, etc. correctly instead of phonetic nonsense.
The seventh mode is different from the voice-driven six: no mic, no Whisper. It's a preview-and-confirm version of the ⌘⌥0 selection-rewriter — you see what got captured, you pick which mode to apply, you review the result, then you decide when to paste it back.
- Highlight text in any app (Mail, Safari, Notes, your IDE, any Electron thing, …).
- Press your configured Office hotkey. There is no default — assign one under Settings → Hotkeys. While the window is open blitzbot also appears in the Dock + ⌘-Tab, so you can jump back to it from anywhere; the Dock icon disappears again when the window closes.
- Blitzbot grabs the selection before stealing focus (AX API first,
⌘Csimulation as fallback) and remembers the source app's bundle ID for the paste-back step. - The window opens pre-filled with the grabbed text. The source-app name appears as a chip above the editor.
- Profile + model switcher — top of the header. Dropdown lists every connection profile; picking one overrides which endpoint this session uses, without mutating your global active profile. The model field pre-fills with the picked profile's preferred model and is editable, so you can e.g. try Opus for this one request and keep Sonnet as your default.
- Pick a mode from the picker row — Business / Plus / Rage / Emoji / Prompt. The default is whatever
Settings → Allgemein → Text umschreiben → Default-Modusis set to (shared with ⌘⌥0). - Tweak the text if needed — the editor is fully editable.
- Hit
⌘↵or click Verarbeiten. LLMRouter sends it through the chosen profile/model with the picked mode's prompt. - The result appears below, and is automatically copied to the clipboard as a safety net.
- Hit
⌘↵again (now labelled In App einfügen) or click the button. Blitzbot closes the Office window, re-activates the source app by bundle ID, and simulates ⌘V so the result lands where your selection was.
- No selection when you press ⌘⌥O? Blitzbot pre-fills from the current clipboard contents (up to the 200 KB text limit). The source-app info is still captured, so paste-back still works.
- Window opened via MenuBar → Office → Öffnen (not the hotkey)? There's no AX selection available because blitzbot already stole focus by the time the popover opens. The editor falls back to clipboard contents; paste-back is unavailable (no source app was captured), the result just stays on the clipboard.
- File drop — drop a text file (
.txt,.md,.json,.csv,.log, plus common code types) into the small dropzone below the editor to replace the input. Hard limit: 200 KB. Binary types are intentionally unsupported. File-origin inputs have no source app, so the result goes to the clipboard only.
- Toggle window: press your Office hotkey a second time to close.
- Dock appearance — while Office is open blitzbot toggles its activation policy to
.regularso you get a Dock icon + ⌘-Tab entry. Closing the window reverts to the menu-bar-only.accessorypolicy. - Auto-copy to clipboard is always on once processing succeeds, regardless of whether you paste-back. That's the "clipboard by default" safety net — even if paste-back fails or you close the window without confirming, the result is there.
- Re-process button appears after a first result so you can try the same input against a different mode/profile/model without re-grabbing.
- Session override, not global — the profile/model picker changes only the current Office session; your global active profile stays put.
- No disk persistence — everything lives in window state, gone on close.
blitzbot ships a pre-send anonymizer for every outbound LLM call. Names, company mentions, places, emails, IPs, URLs, and phone numbers are detected locally and replaced with neutral tokens before any text leaves the app. The LLM's response is then automatically rewritten back into the real terms on the way in, so for the user the flow looks unchanged — but what the provider sees is redacted.
Every path that hits an LLM — all six voice modes, Office Mode, the ⌘⌥0 selection-rewriter, and the recovery retries — goes through the same wrap. One mechanism, consistent guarantee.
Default since v1.2.2: ON. Privacy Mode is active on a fresh install and stays active unless you explicitly turn it off. Your choice is persisted — an explicit off survives app restarts and updates. If you had Privacy Mode explicitly turned off in an earlier version, that setting is preserved.
-
Scan. When a request is about to go out, the input text is analyzed by several local detectors running in parallel:
- Apple's
NLTagger(.nameType)for personal names, organization names, and place names NSDataDetectorfor phone numbers, URLs, and postal addresses (street + number + ZIP + city, any locale Apple's parser knows)- regex for email addresses, IPv4 + IPv6 addresses, MAC addresses, IBAN bank account numbers, and credit-card numbers (with a Luhn-checksum filter to rule out random long numbers)
- plus your own "Immer anonymisieren" list from Settings (case-insensitive, whole-word match) — catches short all-caps abbreviations and internal code names that the NER model misses
- Apple's
-
Substitute. Each detected entity is swapped for a bracketed all-caps token. The full set of kinds:
Kind Placeholder shape Personal name [NAME_1],[NAME_2], …Organization / company [UNTERNEHMEN_1], …Place (city, region) [ORT_1], …Postal address [ADRESSE_1], …Email address [EMAIL_1], …IP address (v4 or v6) [IP_1], …URL [URL_1], …Phone number [TELEFON_1], …IBAN [IBAN_1], …Credit-card number [KREDITKARTE_1], …MAC address [MAC_1], …The indexer grows per kind as new entities appear. The same original value reuses its existing token, so multi-turn context stays consistent — if you mention the same colleague in two dictations, they get the same placeholder both times.
-
Send. The anonymized text goes to the LLM along with a short bilingual instruction appended to the system prompt: "any
[XXXX_N]tokens in the input are placeholders for real entities — keep them verbatim, do not fill them in, do not apply your own anonymization to plain-text words". This prevents the model from either "correcting" the placeholders or inventing new ones of its own. -
Reverse. The model's response is scanned for the placeholders you sent out and they're swapped back into the original values before the text reaches your clipboard, your active app, or the Office preview. Longer placeholders are matched first so
[NAME_10]doesn't get eaten as[NAME_1]+"0]".
Three equivalent ways to flip Privacy Mode — all toggle the same underlying state, all persist.
| Surface | Where |
|---|---|
| Menu-bar popover header | Click the ⚡ icon → the shield pill next to the status indicator is the toggle. Shows live entity count when on. |
| Office window header | Shield pill next to the profile/model pickers. Click opens a popover with the full session mapping ([NAME_1] ↔ Alex Example etc.) and an inline toggle + reset. |
| Recording HUD header | Shield pill between the language badge and the timer. Single click toggles. Compact — no popover on the HUD (nonactivating panel). |
| Settings → Allgemein → Privacy | Full section with the toggle, live entity count, reset button, and the "Immer anonymisieren" editor. |
On a fresh install Privacy Mode starts on. On an existing install your persisted choice wins — if you had it off before, it stays off.
NLTagger is ML-based and sometimes misses short all-caps words or domain-specific code names. The Immer anonymisieren field in Settings → Allgemein → Privacy takes a comma-separated list of terms that are always replaced, case-insensitive, whole-word only. Typical entries: your employer's short name, internal project codenames, your own last name, client code names. Persisted in UserDefaults; stays empty by default.
- No external service ever sees the originals. All detectors are macOS system frameworks, running locally. A privacy feature that phoned home would defeat its own point.
- Mapping stays in memory. Lives on
AppConfig.privacyEngine, never written to UserDefaults, Keychain, or disk. - Reset on toggle-off and quit. Turning the feature off wipes the session mapping; app quit disposes of the engine entirely. No PII dictionary sitting around between runs.
- Log only counts.
~/.blitzbot/logs/blitzbot.logrecordsPrivacy: anonymized name=3 organization=1 …— never the originals themselves. - System prompts pass through unchanged (except for the bilingual instruction added on the fly when Privacy Mode is on). Default prompts contain tool names like
Claude/ChatGPT/CursorthatNLTaggerwould flag; placeholdering them would tank output quality. If your custom prompt overrides contain PII, anonymize it before saving — the feature treats system prompts as canonical app content.
- Green shield pill in the menu bar, Office header, and HUD, with the session's unique-entity count.
- Settings → Allgemein → Privacy: live counter + a Zurücksetzen button.
- Office Mode: click the shield pill to open a popover with the full
placeholder ↔ originalmapping, color-coded by kind (names blue, orgs purple, places teal, emails orange, URLs green, phones indigo, IPs pink). - Log lines per request so you can verify the mapping is firing.
What happens when you've dictated 20 seconds, the Claude call fails halfway through, and you're on a flaky coffee-shop Wi-Fi? Before: the transcript was lost to an auto-hiding error toast. Now: the HUD stays open and offers you a second chance.
Recoverable failures:
- network-level: no internet, timeout, DNS error, TLS failure, host unreachable
- HTTP 401 / 403 (bad / expired key)
- HTTP 5xx (provider outage)
When one of those fires, the HUD switches into recovery mode:
- the transcript is mirrored to the system clipboard immediately (safety net — even if the app crashes now, your text is safe)
- a profile picker appears inline — every connection profile you have, with the just-failed profile marked fehlgeschlagen and disabled
- a 30-second countdown pill starts. If you pick nothing, recovery state is discarded — but the transcript still lives in the clipboard
- click a different profile, hit Erneut senden — the retry reuses the same transcript against the alternate endpoint, without re-transcribing. If the retry also fails, the flow re-enters recovery with the new profile flagged
What's not recoverable (these still go to the old .fehler state): empty transcriptions, "no API key configured", malformed responses, context-length exceeded. Those aren't a profile problem and another endpoint won't fix them.
The retry path is an override — your globally-active profile stays put. Use this when your work VPN proxy is flaky but your direct-Anthropic profile works.
blitzbot can talk to any OpenAI-compatible or Anthropic-compatible LLM endpoint — not just the default Anthropic API. You configure endpoints as profiles, pick one as active, and all LLM calls flow through it.
| Field | Description |
|---|---|
| Name | Display name (e.g. "Direct Anthropic", "Work proxy") |
| Provider | Anthropic / OpenAI / Ollama — controls the request format |
| Base URL | Endpoint root (e.g. https://api.anthropic.com, http://localhost:11434) |
| Auth scheme | x-api-key header / Bearer token / none |
| Active model | Which model to use (picked from a live list or typed manually) |
| API key / token | Stored in the macOS Keychain — never in files |
Open Settings → Profile:
- Quick-switcher chips at the top — click any chip to activate that profile immediately
- New profile button — opens an inline editor
- Scan this Mac — automatically discovers configs in
~/.claude-profiles/,~/.claude/settings.json, and~/.config/claude/and offers to import them as profiles (readsANTHROPIC_BASE_URL,ANTHROPIC_AUTH_TOKEN,ANTHROPIC_API_KEY,ANTHROPIC_MODEL) - Import / Export — JSON format (export never includes secrets)
- Model discovery — inside the profile editor, click Abrufen to fetch the live model list from the endpoint. Click any model row to set it as the active model.
API keys and tokens are stored in the macOS login keychain with an open-access ACL — meaning any application on your Mac can read the item without a per-app confirmation dialog. macOS will never ask for your login password to grant blitzbot access, not even after rebuilds or updates.
Upgrading from v1.0.x / v1.1.0: on the first launch blitzbot rewrites existing keychain items with the open-access ACL (one-time, silent). After that, no prompts ever.
- macOS 13 (Ventura) or newer — Windows and Linux are not supported (see platform note above)
- Apple Silicon (arm64) — Intel not tested
- Homebrew
- About 1.5 GB of free disk for the Whisper model
- Anthropic API key (only for Business / Plus / Rage / Emoji — Normal works without)
- Grab the latest
.zipfrom Releases. - Unzip. Drag
blitzbot.appinto/Applications. - First launch — read this
⚠️ — see First launch on macOS below. - Run the Whisper setup from the repo (see below) — you still need the CLI and model.
blitzbot is ad-hoc signed, not notarized by Apple. That means when you double-click blitzbot.app the first time, macOS Gatekeeper will block it with one of these messages:
- "blitzbot.app" can't be opened because the developer cannot be verified.
- "blitzbot.app" can't be opened because Apple cannot check it for malicious software.
- "blitzbot.app" is damaged and can't be opened. (macOS 15+ Sequoia, if downloaded via browser)
Why: shipping a Gatekeeper-clean macOS app requires enrolling in the Apple Developer Program (99 €/year) and notarizing every release through Apple's service. This project is open source and hobbyist — that cost isn't justified yet. The source is on GitHub, you can read and build it yourself.
How to open it anyway (one-time per install):
- Close the Gatekeeper dialog.
- Finder → Applications → right-click
blitzbot.app→ Open (or Control-click → Open). - A similar dialog appears but now with an Open button. Click Open.
- macOS remembers your decision. From the second launch onward it opens normally.
If the dialog says "is damaged" (macOS 15+), run this once in Terminal:
xattr -dr com.apple.quarantine /Applications/blitzbot.appThat strips the quarantine attribute Safari/Chrome attach to downloaded files. Then double-click normally.
Security note: you're running un-notarized code. That's a real trade-off. Before opening, you can:
- Build from source yourself (
./build-app.sh— see Development) - Inspect the signed binary:
codesign -dv /Applications/blitzbot.app(should saySignature=adhoc) - Diff the zip against what the repo would produce at the tagged commit
The app's behavior is constrained — it records audio, calls whisper.cpp locally, and optionally calls the Anthropic API. No other network calls, no telemetry, no auto-update writes outside /Applications/blitzbot.app itself.
git clone https://github.com/mosandlt/BlitzBot.git
cd BlitzBot
./setup-whisper.sh # installs whisper-cpp via brew + downloads the model
./build-app.sh # compiles + bundles blitzbot.app
open blitzbot.app./setup-whisper.sh does two things:
brew install whisper-cpp— the local CLI- Downloads
ggml-large-v3-turbo.bin(~1.5 GB) into~/.blitzbot/models/— the model used for transcription
The model lives outside the repo so it isn't re-downloaded on every clone.
On first launch blitzbot opens a Setup window that walks you through four checks:
- Microphone — required
- Accessibility — required (for the Cmd+V paste simulation)
- Whisper binary — must exist (opens a Terminal snippet for
brew install whisper-cppif missing) - Whisper model — must exist. If it's missing, click Jetzt laden and blitzbot downloads
ggml-large-v3-turbo.bin(~1.5 GB) directly into~/.blitzbot/models/— no Terminal needed. The download sheet shows live progress and can be cancelled.
Green checkmarks everywhere? Click Continue. Something red? Click the button next to it. If you close the window prematurely, reopen it via Settings → Setup.
Only needed for Business / Plus / Rage / Emoji / Prompt.
Quick setup (direct Anthropic):
- Get a key at https://console.anthropic.com/settings/keys
- In blitzbot: menu bar ⚡ → ⚙ (gear) → Profile → New profile → paste key → Save
- Key is stored in the macOS Keychain. Never in files, never in git.
Custom endpoint or proxy: use the Profile tab to set a different base URL and auth scheme. The scanner (Auf diesem Mac suchen) can auto-import settings from Claude Code config files.
Normal mode runs entirely offline — no key, no cloud calls, nothing. (Optional STT Correction in Settings → General changes this — see below.)
- Put focus into your target app (Notes, Mail, any text field).
- Press your mode hotkey (default
⌘⌥1for Normal). - A HUD appears in the center of the screen with a live timer, animated waveform, mode badge, and status line.
- Speak.
- Press the same hotkey again to stop.
- 1-3 seconds later (Whisper + optional Claude call), text pastes into the target app via simulated Cmd+V.
During recording, a floating panel shows up in the middle of your screen:
- Top-left: X cancel button (aborts recording without pasting) + mode icon + name
- Top-right: language pill (shows detected language, tap to cycle Auto → DE → EN) + cloud icon (tap to toggle STT Correction on/off) + Privacy shield + elapsed time (
mm:ss, monospaced) - Controls row: Pause / Resume button (left) + Auto-Stop countdown clock with draining ring (right, only visible when auto-stop is running)
- Live transcript box: scrollable text area at the top of the HUD — shows your words appearing in real time as you speak, powered by Apple's
SpeechTranscriber(macOS 26+, 16-core ANE). Finalized words are shown in normal weight; the in-flight tail is dimmed. Auto-scrolls to the latest word. Visible as soon as recording starts. Can be disabled in Settings → General. The final pasted text always comes from whisper-cli (higher accuracy); this is visual feedback only. - Middle: full-width real audio waveform — draws actual PCM samples from the mic, scrolling in real-time. Always yellow during recording; grey when idle. A small "Stimme erkannt" badge fades in when voice is detected.
- Silence banner: fades in after 5 seconds of continuous silence with countdown to auto-stop (disappears smoothly when you resume speaking, no layout jumps)
- Status line: Recording… → Pausiert → Transkribiere… → Korrigiere… → Formuliere… → Fertig
- Bottom row: six mode pills (click to switch mode live), Auto-Execute toggle (↵), red Stop button
The HUD does not steal focus — it's an NSPanel with nonactivatingPanel, visible on all Spaces including fullscreen apps. Your target app keeps focus, so the Cmd+V actually pastes where you expect. The mode pills, Pause button, and Stop button work via mouse click because the panel accepts events without activating.
- Stop (red button or hotkey re-press): ends recording and processes + pastes the text
- Cancel (X button, top-left): ends recording and discards everything — nothing gets pasted. Good for accidental presses or drafts you change your mind about.
The menu bar icon reflects current state without requiring you to open the popover:
| Icon | State |
|---|---|
| ⚡ (yellow bolt) | Idle, ready |
| 🔴 red dot + "REC" label | Recording |
| 🟡 waveform | Transcribing or calling Claude |
| 🟢 green checkmark | Done — text pasted |
| 🟠 orange warning triangle | Error (hover/click for detail) |
Click the icon to open the popover with the full mode list, status, and access to Settings.
Seven tabs under ⚙ Settings:
| Tab | What's inside |
|---|---|
| General | Launch-at-Login toggle (SMAppService — blitzbot starts with your Mac), microphone picker (Core Audio device list, falls back to system default if the chosen mic is gone), output language (Auto/DE/EN), auto-stop on silence (toggle + timeout 10s–2min, default 60s), Whisper binary path, Whisper-Modell-Picker (curated list: base / small / medium / large-v3-turbo / large-v3-turbo-Q5 / large-v3 — auto-download on switch, old models purged to save disk; "Benutzerdefiniert" preserves manual paths). Active profile name shown with a quick link to the Profile tab. |
| Profile | Connection profiles — add, edit, delete, import/export, scan for local configs. Quick-switcher chips. Model list per profile. See Connection profiles. |
| Hotkeys | One recorder field per mode. Click, press keys. Defaults shown. Hold-to-Talk toggle (push-to-talk: hold the hotkey to record, release to stop) — alternative to the default toggle behavior. |
| Prompts | Editable system prompt per mode. Leave empty = language-aware default. Add text to either replace or append to the default (toggle per mode). |
| Vocabulary | Proper nouns, product names, jargon, colleagues. Passed to Whisper as --prompt. Improves spelling accuracy dramatically. |
| Setup | Opens the onboarding wizard again. Use when permissions got reset (common after rebuilds). |
| About | Version, update check, GitHub link, license. |
The picker in Settings → Whisper → Modell offers six multilingual variants from ggerganov/whisper.cpp on HuggingFace. All run fully local and fully offline — no audio ever leaves your machine. The differences are speed, disk footprint, and accuracy at the edges (heavy accents, fast speech, low-volume passages, technical jargon).
| Model | Size | Speed¹ | Accuracy DE/EN² | Picks itself for |
|---|---|---|---|---|
base |
148 MB | very fast | drops words on accent / fast speech, weak on German jargon | very small disk, English-only / very simple dictations |
small |
488 MB | fast | acceptable on clear speech, German still wobbly | older MacBook on battery, short clips, low-stakes use |
medium |
1.5 GB | medium | solid DE/EN, near-large on most clips | balanced fallback when Turbo isn't available |
large-v3-turbo |
1.6 GB | fast | excellent DE/EN, very few misreads | default — best speed/quality tradeoff for desktop dictation |
large-v3-turbo-Q5_0 |
~600 MB | fast | minimal quality drop vs Turbo, ⅓ disk footprint | tight on disk, want near-Turbo quality |
large-v3 |
3.1 GB | slow (~2× turbo) | highest accuracy at the edges (accent, jargon, low audio) | heavy accent, technical vocabulary, transcription fidelity > speed |
¹ Wall-clock relative to a 10-second clip on M-series Apple Silicon. large-v3-turbo typically transcribes a 10-s clip in ~1-2 s; large-v3 takes 3-5 s on the same hardware.
² Subjective. None of these are "wrong" — they trade off how often they nail an unusual word vs. how fast you get the result.
Switching: pick a different entry from the dropdown. If the chosen file isn't already in ~/.blitzbot/models/, blitzbot opens the download sheet automatically. After a successful switch, all other ggml-*.bin files in that folder are deleted so disk doesn't grow with every experiment. Manual paths (Benutzerdefiniert) are exempt from auto-cleanup.
Quantization note: the Q5_0 variant uses 5-bit weight quantization. The model is mathematically smaller and a touch lossier, but on a desktop dictation workload (clean mic, ~10-30 second clips, common vocabulary) the output is rarely distinguishable from large-v3-turbo proper. It's the right pick when 1 GB of disk matters more than the last few percent of edge-case accuracy.
Don't pick base or small for German. Whisper's smaller multilingual variants are noticeably weaker on German than English, and umlauts / compound words get butchered. medium is the realistic floor for German dictation.
microphone
↓
AVAudioEngine ──► /tmp/blitzbot-<uuid>.wav (ephemeral, deleted after transcription)
↓
whisper-cli (local, offline, no network)
↓
text
↓
optional STT correction pass (Settings → General → "STT-Korrektur"):
└─ LLM call to fix Whisper mis-transcriptions, dialect mix-ups, missing punctuation
→ corrected text (or original if LLM unavailable — silent fallback)
↓
mode router (for non-Normal modes, LLM provider is configurable: Claude / OpenAI / Ollama):
├─ Normal → text directly (no cloud call, regardless of provider)
├─ Business → LLM call with business prompt
├─ Plus → LLM call with light-touch polish prompt
├─ Rage → LLM call with de-escalation prompt
├─ Emoji → LLM call with emoji-insertion prompt
├─ Prompt → LLM call that turns a loose idea into a tool-agnostic prompt
└─ Office → LLM call with structured-summary prompt (bypasses voice path entirely;
input is typed text or a dropped file, output lands in a preview window
instead of being auto-pasted)
↓
NSPasteboard.general (writes the result)
↓
CGEvent Cmd+V simulation (120 ms delay, cgAnnotatedSessionEventTap)
↓
text lands in whatever app has keyboard focus
Privacy guarantees:
- Audio never leaves your machine. Not in any mode. Transcription is 100% local via whisper.cpp.
.wavfiles are temporary. Created in/tmp/, deleted immediately after Whisper finishes (inside adeferblock in the transcriber).- Normal mode makes zero network calls by default. No telemetry, no analytics, no phone-home. (Exception: if you enable STT Correction in Settings → General, Normal mode also sends the transcript to your configured LLM provider for a dialect-fix pass before pasting.)
- Business / Plus / Rage / Emoji / Prompt send exactly one HTTPS request to
api.anthropic.com, containing just the transcribed text + the mode's system prompt. That's the only thing that ever leaves your machine. - The API key lives in the macOS Keychain, not in UserDefaults, not on disk in plain text.
- Transcripts aren't persisted. The dev log at
~/.blitzbot/logs/blitzbot.loglogs transcript length only — the content itself is not written to disk.
| Mode | Cost per dictation (rough) |
|---|---|
| Normal | $0 — fully local (+ ~$0.0003 if STT Correction is enabled) |
| Business | ~$0.0006 — 200 input tokens × Sonnet pricing |
| Plus | ~$0.0006 |
| Rage | ~$0.0006 |
| Emoji | ~$0.0006 |
Pricing source: Anthropic pricing — update yourself if rates change. In practice dictating all day long costs a few cents.
blitzbot touches three macOS permission systems:
- Microphone — so it can record you. Granted via the first
AVCaptureDevice.requestAccesscall; macOS remembers per user account. - Accessibility — so it can simulate the Cmd+V that pastes into the active app. Granted in System Settings → Privacy & Security → Accessibility.
- Input Monitoring / Post Events — implicitly covered by Accessibility on recent macOS; on some older versions you may also need to toggle Input Monitoring.
Short answer: ad-hoc code signing.
Long answer: macOS TCC (the Transparency, Consent, and Control system) binds permissions to a specific code signature hash (cdhash). Every time you run codesign --force --sign - with an ad-hoc identity, the cdhash changes. TCC then doesn't recognize the app as the same entity that was previously granted permission, and revokes it.
The microphone permission survives rebuilds because it's stored in the user TCC database and keyed slightly differently (by bundle ID). Accessibility and Input Monitoring live in the system TCC database and are stricter.
Create a self-signed certificate once, then always sign with it. Permissions stick.
- Open Keychain Access.
- Menu: Certificate Assistant → Create a Certificate…
- Name:
blitzbot-dev - Identity Type: Self Signed Root
- Certificate Type: Code Signing
- Click Create.
- After creation, double-click the cert in Keychain Access → Trust → When using this certificate: Always Trust (specifically mark Code Signing as Always Trust).
Then sign with it:
./build-app.sh --sign blitzbot-devPermissions now persist across rebuilds. Until you do this step, you'll need to re-grant Accessibility after every build — the Setup tab makes this a two-click operation, but it's annoying.
| Symptom | Likely cause | Fix |
|---|---|---|
| Menu bar icon appears, but hotkey does nothing | Accessibility permission revoked | Settings → Setup → Open in Settings → re-enable blitzbot |
| Text ends up in clipboard but doesn't paste | Accessibility permission revoked | Same as above |
| "No API key" warning at bottom of popover | No Anthropic key saved | Settings → General → paste key → Save |
whisper-cli not found |
Whisper not installed, or path changed | Settings → General → Whisper binary → set correct path, or run ./setup-whisper.sh again |
| Transcription is garbage or wrong language | Whisper language assumption (default: de) | Edit WhisperTranscriber.swift:6 — language: String = "de" → change to your language code |
| Business / Plus / Rage output is empty | Claude API call failed (e.g. invalid key, rate limit) | Check ~/.blitzbot/logs/blitzbot.log for an ERROR: line |
| Icon is always black / doesn't render | Rebuilt icon with the dev script, ad-hoc sign didn't refresh LaunchServices | Run touch blitzbot.app then open blitzbot.app — forces LaunchServices refresh |
| HUD doesn't appear | Processor didn't transition to .aufnahme state |
Log check — either the audio engine failed to start (mic permission?) or the mode toggle didn't fire |
tail -f ~/.blitzbot/logs/blitzbot.logEvery state transition is logged: hotkey registration, recording start/stop with byte count, transcript preview, paste confirmation, and any errors. This is your first debugging tool.
- Swift 5.9+ with macOS 13+ as minimum target
- Swift Package Manager — no Xcode project file, just
Package.swift - Xcode or any Swift-aware editor
- External dependency:
KeyboardShortcutsby Sindre Sorhus, pinned<1.15.0(the 2.x line uses#Previewmacros that break non-Xcodeswift build)
Build artifacts live in ~/Downloads/blitzbot-build/, never inside the project directory — the project may live in a synced folder (Nextcloud, iCloud, Dropbox) and the .build/ folder gets hundreds of megabytes.
# Default: sign with local blitzbot-dev cert
# (permissions + keychain stay intact across rebuilds on your own machine)
./build-app.sh
# Explicit identity
./build-app.sh --sign <identity>
# Release build: ad-hoc sign + zip for GitHub release
# (portable, but end users will hit Gatekeeper on first launch)
./build-app.sh --release
# Launch
open ~/Downloads/blitzbot-build/blitzbot.app
# Regenerate the app icon (writes blitzbot.iconset/ + updates the committed AppIcon.icns)
swift tools/make-icon.swift
iconutil -c icns blitzbot.iconset -o blitzbot.app/Contents/Resources/AppIcon.icns
# Follow logs
tail -f ~/.blitzbot/logs/blitzbot.log
# Run Whisper on a file directly (debugging)
whisper-cli -m ~/.blitzbot/models/ggml-large-v3-turbo.bin -f test.wav -l de -nt -otxt -of test| Scenario | Command | Cert | Who can open the app |
|---|---|---|---|
| Local dev (your own Mac) | ./build-app.sh |
blitzbot-dev (self-signed, in your login keychain) |
Only you — signature tied to the private key on your machine |
| GitHub Release | ./build-app.sh --release |
Ad-hoc (no cert) | Anyone, but first launch triggers Gatekeeper (see First launch) |
| Custom identity | ./build-app.sh --sign <name> |
whatever you pass | Depends on the cert |
The local blitzbot-dev cert is regenerated via the CLI block in CLAUDE.md if it ever gets lost. No Keychain Access GUI needed. The cert stays on your machine — never bundled into releases.
Shipping a Gatekeeper-clean release requires the Apple Developer Program (99 €/year) and notarization. Not done here. See First launch for what end users see.
Sources/blitzbot/
blitzbotApp.swift @main + AppDelegate + MenuBarExtra + Windows (Settings / Setup / Office)
AppConfig.swift UserDefaults, Keychain wrapper, vocabulary, per-mode prompts
AppInfo.swift Version constants and repo URL
Mode.swift enum (Normal…Prompt…Office) + display names + default prompts
HotkeyManager.swift KeyboardShortcuts integration per mode + rewriteSelection + toggleOffice
ModeProcessor.swift state machine: toggle → record → transcribe → formulate → paste.
Also owns the inline-recovery flow (RecoveryContext, 30 s timer,
profile-switch retry) for recoverable LLM errors.
AudioRecorder.swift AVAudioEngine setup + rolling PCM sample buffer + RMS level publishing
WhisperTranscriber.swift subprocess wrapper around whisper-cli
LLMRouter.swift routes LLM calls through the active profile or legacy fallback;
also provides a profile-override overload used by recovery retries
LLMError.swift structured error type (connectionFailed / authFailed / serverError /
other). `isRecoverable` drives the inline-recovery UI.
AnthropicClient.swift Claude API request/response (supports custom base URL + auth schemes);
throws `LLMError` for recoverable failures
OpenAIClient.swift OpenAI-compatible API client (same error contract)
OllamaClient.swift Ollama local-LLM client (same error contract)
Paster.swift NSPasteboard + CGEvent Cmd+V simulation
KeychainStore.swift API key read/write/delete — open-access ACL, no prompts ever
KeychainPreWarmer.swift one-time ACL migration at first launch → silent forever after
ConnectionProfile.swift profile model (provider, baseURL, authScheme, model, Keychain slot)
ProfileStore.swift @ObservedObject store — CRUD, UserDefaults persistence, Keychain I/O
ProfileScanner.swift scans ~/.claude-profiles/, ~/.claude/settings.json for importable configs
ModelDiscovery.swift fetches live model list from Anthropic /v1/models, OpenAI, Ollama
Log.swift simple append-only log at ~/.blitzbot/logs/blitzbot.log
Permissions.swift TCC status checker for mic/accessibility/whisper
PermissionsView.swift onboarding wizard UI
MenuBarView.swift popover content (header, mode list, footer); Office row opens its window
SettingsView.swift TabView (General / Profile / Hotkeys / Prompts / Vocabulary / Setup / About)
ProfilesView.swift Profile tab UI — list, editor, scanner, quick-switcher
RecordingHUD.swift NSPanel + SwiftUI content for the floating recording HUD.
Renders the inline recovery UI (profile picker + countdown) when
`ModeProcessor.Status` is `.recovery`.
LiveTranscriber.swift Apple SpeechTranscriber (macOS 26+) wrapper — feeds PCM buffers from
the AVAudioEngine tap into SpeechAnalyzer and streams LivePartial
(confirmed + volatile text) back to the HUD in real time.
OfficeView.swift Office Mode window — dropzone, text editor, Verarbeiten button,
result preview, copy-to-clipboard. Calls LLMRouter directly.
SelectionRewriter.swift ⌘⌥0 hotkey — reads AX selection, rewrites via LLM, pastes back
Updater.swift GitHub Releases API check + download + install
LaunchAtLoginManager.swift SMAppService.mainApp wrapper — register/unregister as Login Item,
surfaces .requiresApproval state so the toggle can't lie
ModelDownloader.swift streams ggml-large-v3-turbo.bin from HuggingFace into
~/.blitzbot/models/ with progress, cancel, and GGUF-magic verify
Claude-Code skills shipped in .claude/skills/:
build-deploy— rebuild + deploy to/Applications, auto-runstools/smoke-test.shfor log verificationrelease— guided release flow (version bump → build → tag → pause for user go → push +gh release createwith bilingual notes)usage-report— on-demand dictation stats viatools/usage-report.py(per-mode counts, P50/P95 latency, language split)pre-push-scan— mechanical secrets/PII scan before every pushdev-cert-regen— rebuildblitzbot-devcode-signing cert if lost
Scripts in tools/:
smoke-test.sh— checks the log for the four startup markers after deployusage-report.py— parses~/.blitzbot/logs/blitzbot.logand prints a text or JSON summarymake-icon.swift— regenerates the app icon
See CLAUDE.md for the full set of rules this project follows. Highlights:
- Two-prompt rule: before any non-trivial feature, run a critical-analysis pass first, then implement. No "just code it up".
- Log via
Log.write(...), neverprint, neverFileHandle.standardError.write— the persistent log survives app restarts and is your main debugging tool. - Lifecycle via
NSApplicationDelegate(wired with@NSApplicationDelegateAdaptor). SwiftUI.onAppearonMenuBarExtralabels is not reliable as a startup hook. - Floating panels must not steal focus — always
NSPanelwith.nonactivatingPanel+ignoresMouseEvents. - No new external dependencies without discussion. Every package added is supply-chain risk and bundle-size tax.
┌──────────────────────────────────────────────────────────────────────┐
│ user presses ⌘⌥2 (Business) │
└──────────────────────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────────────────┐
│ KeyboardShortcuts → HotkeyManager.onTrigger → ModeProcessor.toggle │
└──────────────────────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────────────────┐
│ ModeProcessor state machine │
│ bereit → aufnahme → transkribiert → formuliert → fertig → bereit │
│ │
│ In aufnahme: AudioRecorder running + @Published elapsed timer │
│ + RMS level stream into HUD │
│ In transkribiert: WhisperTranscriber subprocess │
│ In formuliert: AnthropicClient request │
│ In fertig: Paster.pasteText() + 1.5s cooldown │
└──────────────────────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────────────────┐
│ observers │
│ MenuBarLabel (icon + REC badge) │
│ MenuBarView (popover content) │
│ RecordingHUDController (shows NSPanel during active recording) │
└──────────────────────────────────────────────────────────────────────┘
Everything is driven by @Published fields on ModeProcessor and AudioRecorder. SwiftUI views subscribe via @EnvironmentObject. No shared mutable state outside these observables.
PRs welcome. A few principles:
- Small, focused commits. No "wip", no "fix stuff".
- Keep the tagline pattern. New mode? Tagline must read "X in. Y out." in both English and German.
- Don't add dependencies lightly. The current dep list is
KeyboardShortcutsand that's it. Be prepared to justify additions. - German + English for UI strings. See
blitzbot.app/Contents/Resources/en.lproj/Localizable.stringsand theString(localized:)calls in source. - Before pushing: run the secrets scan in
CLAUDE.md(section Sicherheit & Privatsphäre). No.env, no API keys, no personal identifiers.
Areas where help is especially welcome:
- Push-to-talk as an alternative to toggle (hold hotkey while speaking)
- Intel x86_64 build + universal binary
- Notarization pipeline for the release workflow
- Additional languages in
Localizable.strings - More modes (e.g. Translate — dictate in one language, get output in another)
blitzbot itself will stay macOS-only — the tight OS integration is a feature, not a bug, and porting would mean a parallel codebase that drifts over time.
If you want to build a sibling (Windows, Linux, iOS, Android, web) the approach is:
- New repo (or new top-level directory in this repo, e.g.
blitzbot-win/,blitzbot-ios/) - Share the things that are genuinely portable: the Claude system prompts (
Mode.swift), the Whisper invocation, the API request shape - Open an issue first to align on scope so we don't end up with multiple half-finished ports
I personally won't maintain ports — I can't test on OSes I don't use. But I'll happily link to and endorse a well-built sibling project from this README.
- Cross-platform single codebase: Tauri (Rust + webview) → one binary per OS, small, native-feeling
- Native: C# / WinUI 3 with
RegisterHotKey+SendInput+SetClipboardData
- Native: GTK4 / Qt with
xdotoolor Wayland-specific hotkey handling, which is unfortunately still fragmented (per-compositor APIs, no unified global-hotkey story yet)
Someone asked. We looked into it. Here's the honest take so the next person doesn't have to start from zero.
The core blitzbot UX does not port to iOS. Apple forbids any app from simulating keyboard input into another app. There are also no global hotkeys outside the Shortcuts app. That means "press a key, speak, text lands in whatever app has focus" — the whole point on macOS — is architecturally impossible on iOS. A Windows/Linux port is a rewrite of the OS layer; an iOS "port" is a different app category.
The realistic iOS product is a standalone dictation app — not a system-wide paste engine. Call it blitzbot Notes or similar. User flow:
- Open app (or trigger via Shortcuts / Share Extension / Home-Screen widget)
- Tap mic, speak, tap again
- Pick a mode (Normal / Business / Plus / Rage / Emoji) — same prompts as macOS
- Result shows in an editable text view
- Copy to clipboard, share-sheet into the target app, or save as a local draft
Suggested MVP architecture:
BlitzbotApp (SwiftUI, iOS 17+)
├── RecordView big mic button, mode picker, live waveform
├── TranscriptView editable text, re-process with different mode, copy/share/save
├── DraftsView local history of the last ~50 transcripts (Core Data)
├── SettingsView API key, mode prompts, vocabulary, Whisper model size
└── ShortcutsIntents App Intents for Shortcuts.app integration + Share Extension
Top risks to validate before committing (do these first, in this order):
- Whisper on-device performance. Use WhisperKit or whisper.cpp iOS bindings. Measure latency + heat + battery with three model sizes (
tiny~75 MB,base~150 MB,small~480 MB). Forgetlarge-v3-turbo(1.5 GB) on anything below an iPhone 15 Pro. If this isn't fast + cool + battery-friendly, stop. - Background processing. iOS kills background tasks after ~30 s. Long dictations plus Whisper runtime might get killed mid-transcription. Requires
UIBackgroundModes → audioand careful state restoration. - Thermal throttling. Consecutive long transcriptions warm up the device and the OS starts throttling CPU/GPU. Measure under "normal user" conditions (back-to-back dictations, not a single clean bench).
- Clipboard tax. Since iOS 14,
UIPasteboard.generalreads are surfaced to the user with a toast. Writes are fine, but a flow that reads clipboard contents would hurt trust. - Shortcuts-app integration UX. Technically nice, but onboarding is painful — users have to create a Shortcut, grant permissions, assign an activation method. Most won't bother.
- Custom Keyboard Extension. Theoretical path to "type with your voice everywhere". Blocker: custom keyboards without Full Access cannot hit the network → no Claude call. With Full Access, users get a scary permissions dialog that tanks adoption.
- App Store review. LLM-based text apps are getting stricter content-safety review. The Rage mode ("de-escalate angry text") could flag on the wrong reviewer.
- Model distribution size. Bundled Whisper model bloats the
.ipa. Downloading on first launch is cleaner but needs a resume-capable downloader + integrity check. - API key storage. iOS Keychain works fine, but sync across macOS/iOS via iCloud Keychain is non-trivial to do cleanly.
- Maintenance cost. Two platforms = doubled prompt-sync effort, doubled UI to keep aligned. Without a clear user-value-add beyond the macOS app, this is scope creep.
What's safely in scope for an MVP:
- Dictation with Whisper on-device + mode picker
- Copy-to-clipboard + Share-sheet
- Vocabulary list synced from macOS via iCloud KV store
- TestFlight beta distribution
What's out of scope for an MVP:
- Simulating "paste into active app" — not possible
- Custom Keyboard — too much UX friction for v1
- Always-on background dictation — OS will kill it
Suggested first step: one-day spike. Before planning the whole app, build a throwaway iOS app that does only Whisper on-device with three model sizes and logs latency + temperature + battery delta. If the numbers are bad on realistic hardware, iOS is a bad investment and we should document that instead of shipping a slow app. If the numbers are good, greenlight the MVP planning with actual data.
If you want to take this on: fork or open an issue, propose the spike, and we'll link your repo from here.
- Apple notarization (removes first-launch Gatekeeper workaround for end users)
- Universal binary (arm64 + x86_64)
- Shortcuts.app integration (AppIntents — "Dictate now in mode X")
Got other ideas? Open an issue.
- Voice Isolation no longer leaves the system audio ducked. When VPIO (
AVAudioInputNode.setVoiceProcessingEnabled(true)) is active, macOS lowers system output volume to prevent feedback — the same mechanism FaceTime/Zoom use. Before this fix,AudioRecorder.stop()only removed the tap and stopped the engine; the voice-processing flag stayed armed on the input node, so the OS kept ducking output until app-quit. Re-toggling Voice Isolation didn't help because the early-return guard sawisVoiceProcessingEnabled == trueand skipped reinitialisation. Now stop() explicitly disables voice processing, callsengine.reset()to release the audio unit, and clears the cached flag — system volume returns to normal the moment recording ends.
- Live transcription now works with Voice Isolation. When VPIO is enabled, the input node delivers Float32 multi-channel buffers; Apple's
SpeechTranscriberexpects Int16 mono, so live text was silently empty. The buffer-tap path now extracts channel 0, converts Float32→Int16, and feeds SpeechTranscriber correctly. - Waveform stutter fixed under VPIO. Buffer size at 16 kHz dropped from 4096 frames (256 ms / ~4 callbacks per second) to 1024 frames (64 ms / ~15 callbacks per second), so the waveform animates smoothly instead of stepping in chunks.
- STT Correction — optional LLM pass after Whisper (Settings → General → Transkription → STT-Korrektur). Fixes mis-transcriptions, dialect mix-ups, and missing punctuation before the text reaches any mode processing. Toggle on/off directly from the HUD cloud icon.
- Bavarian dialect fix — when STT Correction is on and Output Language is Auto, blitzbot now forces
-l deto Whisper so Bavarian (and other German dialects) are no longer mis-detected as Nordic languages (Icelandic, Danish). The correction prompt knows Bavarian words like ned, hoid, des, wia and maps them back to standard German. - Live language detection — the language badge in the HUD now updates in real-time during recording. On Auto, it shows the actually detected language (DE/EN) instead of "AUTO" as soon as enough speech is recognised — both from Apple's SpeechTranscriber (macOS 26+ ANE) and from Whisper after processing.
- HUD quick-controls — two new interactive elements in the HUD header: a language pill (tap to cycle Auto → DE → EN → Auto without opening Settings) and a cloud icon (tap to toggle STT Correction on/off mid-session). Both changes are saved immediately.
- Live transcription in the HUD — while you speak, your words appear in real time in a scrollable transcript box at the top of the recording HUD. Powered by Apple's
SpeechTranscriber(macOS 26+, requires 16-core ANE / Apple Silicon M-series). Enable or disable in Settings → General → Live-Transkription. The final pasted text always comes from whisper-cli (better accuracy); the live view is visual feedback only so you're never staring at a blank screen. - HUD layout fixes — transcript box height 44→80 pt (fits ~5-6 lines), auto-scrolls to the latest word as you speak, visible immediately when recording starts (no jump on first word). Panel height 300→380 pt so the mode pills at the bottom are no longer clipped.
- Translate mode (#7) —
⌘⌥7. Auto-flip translator: dictate in German → get English back, dictate in English → get German back. Whisper detects the source language; Claude translates into the other one preserving register and tone. Names, brands, code identifiers stay untouched. Single prompt handles both directions, no extra setting. - Hold-to-Talk — Settings → Hotkeys gets a global toggle. When on, mode hotkeys behave as push-to-talk: keyDown starts recording, keyUp stops. Auto-repeat keyDowns are suppressed. Default stays toggle (the v1.0 behavior).
- Whisper-Modell-Picker — Settings → Whisper → Modell shows a curated list:
base,small,medium,large-v3-turbo(Empfohlen),large-v3-turbo-Q5,large-v3. Switching auto-downloads the chosen variant on demand and purges otherggml-*.binfiles in~/.blitzbot/models/to keep disk in check. "Benutzerdefiniert" preserves the manual-path escape hatch. - Multi-Mic-Selector — Settings → Mikrofon. Core Audio device enumerator with a picker; stored as the device UID so reconnects survive USB re-enumeration. If the chosen mic is gone at recording start, AudioRecorder logs and falls back to system default — recording never fails because of a missing mic.
- HUD progress feedback — the 2-25 s wait between REC stop and PASTE used to show a static "Transkribiere…" / "Formuliere…" line. Now shows a spinner, a live tenths-of-a-second counter, and the active provider name during the LLM step ("Formuliere via Anthropic … 4.2s"). No more "is it stuck?" feeling on long Opus calls.
- Whisper decoding stability —
WhisperTranscribernow passes--temperature 0,--no-fallback,--suppress-nst, and explicit beam/best-of pinning. Kills whisper.cpp's temperature-fallback that paraphrased low-confidence segments into plausible-but-wrong sentences. Tradeoff: noisy clips lose the safety net and may produce a few garbled words instead of a smooth rewrite. For dictation where the user reviews output, literal beats fluent. - Recovery: WAV header repair —
AVAudioFileonly writes the RIFF + data chunk sizes on close, so a recording killed mid-stream leaves a file with audio bytes but a 0-byte data chunk. Whisper then read 0 frames and Recovery logged "transcript empty". The orphan-recovery path now patches the chunk-size fields from the actual file size before transcribing, so killed recordings come back intact. - Startup-timing log marker — added
applicationWillFinishLaunchingto the log so a future slow-startup incident (a 33-s gap was observed once) can be bisected.
- Launch-at-Login toggle under Settings → General. Uses
SMAppService.mainAppso blitzbot starts with your Mac and stays as the menu-bar icon (no Dock icon). The toggle reflects the real system state, including macOS's.requiresApprovalcase with a deep-link button to open System Settings → General → Login Items. - In-app Whisper model download. The first-launch Setup window now has a Jetzt laden button next to the Whisper-Modell row. It streams
ggml-large-v3-turbo.bin(~1.5 GB) from HuggingFace directly into~/.blitzbot/models/with live progress, cancel, and a GGUF-magic-byte sanity check. No more Terminal detour for new users;setup-whisper.shstays available for power users. - Tooling:
tools/smoke-test.sh(verifies four startup markers + panic patterns in the log after every deploy) andtools/usage-report.py(per-mode dictation counts, P50/P95 latency, language split, incomplete sessions — on-demand, no cron). New Claude-Code skillsrelease(guided release pipeline with user-go before push) andusage-report. Existingbuild-deployskill now auto-runs the smoke test.
- Prompt caching for Anthropic direct API.
AnthropicClientnow sends the system prompt withcache_control: { type: ephemeral }when the active profile hitsapi.anthropic.comwithx-api-keyauth. Proxies and custom-baseURL profiles send a plain string as before — they may rejectcache_controlblocks. Effect: ~90% reduction in input tokens for repeated calls within the 5-minute TTL window; cost drop is visible in the Anthropic usage dashboard on any workload with repeated mode use. - Fix: Prompt mode effort restored to
xhigh. During v1.3.4 optimization work theaiCommandeffort level was inadvertently lowered fromxhightohigh. The Prompt mode system prompt is the most demanding in the codebase: multi-signal update-vs-new-project detection, 5-part structured output as flowing text, and eight simultaneous constraint rules (no meta-sentence, no code generation, no preamble, imperative voice, …). Athighbudget Opus 4.7 drops the lower-priority constraints and occasionally generates code instead of a prompt. Reverted toxhigh. - Anti-inflation hint for Business mode. Added explicit length constraint to the Business DE+EN system prompts: output should be roughly as long as the dictation, not longer. Prevents Opus 4.7's tendency to expand short inputs with boilerplate padding.
- Shape-based fixture tests.
tests/mode-fixtures/— input + expected-criteria files for all five LLM modes (Business, Plus, Rage, Emoji, Prompt). Run with./tests/mode-fixtures/run-fixtures.sh. Criteria:contains,not-contains,max-length,min-length,no-preamble,no-markdown-headings,max-emojis.
- Fix: Settings → Hotkeys crash. The SPM-generated
KeyboardShortcuts_KeyboardShortcuts.bundlewas never copied into the.appbybuild-app.sh. At runtimeBundle.moduletripped_assertionFailurethe instantSettings → Hotkeystried to instantiate a shortcut recorder, and the app silently died with no error dialog. Build script now copies every*.bundleemitted next to the release binary intoContents/Resources/. - Fix: profile list silently emptied after v1.3.2 upgrade.
ProfileStore.loadFromDiskdecoded the whole profiles array withtry?; a single profile with the removedappleIntelligenceprovider made every remaining profile disappear from the UI. Loader now falls back to per-element decoding, keeps everything it can parse, drops only the unknown entries, rewrites the plist cleanly, and repairsactiveProfileIDif it pointed at a dropped profile. Keys stay in the Keychain, nothing to re-enter.
- Apple Intelligence integration removed. v1.3.0 added it as a provider and v1.3.1 made the Privacy wrap skip it. Empirical testing with five real prompts across Plus / Business / Rage / Emoji / Prompt / Office showed the on-device ~3B foundation model fails every LLM-backed mode: it ignores the system prompt, hallucinates fluently on absurd inputs, and loops on repetitive output. The provider was a marketing bullet without user-visible benefit, so it's out.
AppleIntelligenceClient.swiftdeleted,LLMProvider.appleIntelligencecase removed, all UI surfaces (Profile editor, icon, color, availability badge, quality hint table) removed.- If you want a larger local model: use Ollama (already supported since v1.0.7).
ollama pull qwen2.5:14b(orqwen2.5:32b/llama3.3:70bif your Mac has the RAM) runs a real large model on-device with quality close to cloud options. Point a profile athttp://localhost:11434and pick the model you pulled. - If Apple ships a larger on-device model in a future macOS release and exposes it via
FoundationModels, we'll reconsider.
- Privacy Mode auto-skips for Ollama (the useful half of v1.3.1, kept). When the active profile's provider is Ollama, the pre-send anonymizer short-circuits: raw text goes straight to the local model, reverse-mapping step is bypassed. Skipping for Ollama is a pure win — no privacy loss (data stays on your Mac), better output quality (no over-anonymization of code identifiers like
meineFunktion,my_var,file.swift).- Log line
Privacy: skipped (local provider: ollama)in~/.blitzbot/logs/blitzbot.logmakes the bypass observable. - UI toggle state unchanged — the shield pill still reflects whether Privacy is armed; it just no-ops for Ollama requests.
- Log line
- Migration: if you had a v1.3.0 Apple-Intelligence profile saved, it stays in UserDefaults but is no longer selectable in the provider picker. Delete it manually in Settings → Profile. Nothing else changes.
- Opus 4.7 per-mode effort hints. When the active model is
claude-opus-4-7, each mode now sends a matchingoutput_config.effortvalue (low / medium / high / xhigh) so the model invests reasoning proportionally to the task: Plus/Emoji = low, Rage = medium, Business/Office = high, Prompt = xhigh. Threading goes throughLLMRouter→AnthropicClientwithout affecting Sonnet/Haiku or other providers. No UI change — the hint is derived automatically from the mode. - Docs cleanup:
CLAUDE.mdrefreshed — stale version pin replaced (was still v1.1.0), mode list extended to the 7th (Office), file overview synced with the full current Swift layout (addsLLMRouter,LLMError,ConnectionProfile,ProfileStore,ProfileScanner,ModelDiscovery,OpenAIClient,OllamaClient,PrivacyEngine,OfficeView,ProfilesView,KeychainPreWarmer), architecture diagram now shows the Privacy-wrap + inline-recovery branch, Settings-Tab count corrected (7 not 6), staleLSUIElement=YESclaim replaced with the correct activation-policy-set-programmatically behavior. Release history inCLAUDE.mdbrought up to date with v1.2.0–v1.2.3. - Build-cache hygiene: removed stray
.build/+ old v1.0.0 release zip that had accumulated inside the Nextcloud-synced project folder (rule fromCLAUDE.md§1).
-
Expanded PII coverage. Privacy Mode now additionally detects and anonymizes:
- Postal addresses — full street + house number + ZIP + city, via
NSDataDetector(.address). Replaced with[ADRESSE_1],[ADRESSE_2], … - IBAN bank account numbers — two-letter country code + two check digits + 11–30 alphanumeric, with or without whitespace blocks. Replaced with
[IBAN_1], … - Credit-card numbers — 13–19 digit sequences that pass a Luhn checksum check. The checksum filter rules out random long numbers (order IDs, reference numbers) that happen to match the digit count, so false-positives stay low. Replaced with
[KREDITKARTE_1], … - MAC addresses — colon- or dash-separated six-group hex (
aa:bb:cc:dd:ee:ff). Replaced with[MAC_1], … - IPv6 — the full eight-group form is now caught alongside IPv4 under the same
[IP_n]placeholder.
All detection remains 100% local (Apple system frameworks + regex + a Luhn helper). No new external dependencies. The mapping popover and settings list color-code the new kinds.
- Postal addresses — full street + house number + ZIP + city, via
-
Privacy Mode now defaults to ON. New installs have the pre-send anonymizer active out of the box; existing installs that previously had Privacy explicitly turned off keep their setting (persisted value wins over the registered default). Switches in Settings → Allgemein → Privacy, the menu-bar popover, the Office window header, or the recording HUD all flip the same state.
-
What the anonymizer does. Before a request goes to an LLM, the outbound text is scanned locally (no external service) and entities in the following categories are replaced with stable placeholders:
Category Detector Placeholder shape Personal names Apple NLTagger(.nameType)[NAME_1],[NAME_2], …Organizations / companies NLTagger+ user-supplied Immer anonymisieren list[UNTERNEHMEN_1], …Place names NLTagger[ORT_1], …Email addresses regex [EMAIL_1], …IPv4 addresses regex [IP_1], …URLs NSDataDetector(.link)[URL_1], …Phone numbers NSDataDetector(.phoneNumber)[TELEFON_1], …The model's response is rewritten back into the real terms on the way in, so your clipboard / pasted text / Office preview always shows the original values — only what leaves the machine is redacted. The mapping lives only in memory, is wiped on toggle-off and on app quit, and is never written to disk.
-
System-prompt hint rewritten to stop the LLM from hallucinating placeholders. The previous version listed example placeholder shapes which led some models to re-anonymize plain-text names on their own (producing placeholders that had no reverse mapping). The new instruction is token-shape-based only and explicitly tells the model not to introduce placeholders of its own.
-
"Immer anonymisieren" custom term list (Settings → Allgemein → Privacy). Terms that are always anonymized, even if
NLTaggermisses them — useful for short all-caps company abbreviations, internal project codenames, or your own last name. Case-insensitive, whole-word match. Persisted separately from the session mapping. List-based UX: each term gets its own row with a minus button; a dedicated input + plus button (Return also adds) for new entries. -
Session mapping inline in Settings. The Privacy section now shows the current
placeholder ↔ originaltable directly in Settings → Allgemein → Privacy (color-coded by kind), in addition to the existing Office-Mode popover. Makes it easy to audit what the engine has captured without jumping into another window. -
Menu-bar shield toggle (small improvement). The privacy shield is now also in the menu-bar popover header next to the status indicator — gives a quick visual confirmation of the active state without having to open Office Mode, a recording HUD, or Settings. Same toggle, same state.
- Privacy Mode: opt-in local anonymization for every outbound LLM call. Personal names, organizations, places (via
NLTagger), emails / IPv4 (regex), URLs + phone numbers (NSDataDetector) are replaced with stable placeholders like[NAME_1]before leaving the app, and the model's response is rewritten back into your real terms before it hits the clipboard or gets pasted. Mapping lives in memory only, wiped on toggle-off and app quit — no persisted PII database. Toggle in Settings → Allgemein or via the shield pill in the Office window header and the recording HUD (same state, live entity count). - Model dropdown in Office. The earlier text field is replaced with a live-fetched list of models from the current profile's endpoint (via the existing
ModelDiscovery). Picking one overrides the model for this session only. Shows a Profil-Standard entry to reset. - Office hotkey now opt-in. The earlier ⌘⌥O default collided with other apps; a one-time migration strips it for users still on the default. Set your own combo in Settings → Hotkeys.
- Per-session profile switcher in the Office header. Pick any connection profile from a dropdown for this one request, without mutating the globally active profile.
- Dock visibility while Office is open —
LSUIElementremoved from Info.plist and.accessoryset programmatically inapplicationWillFinishLaunching. Toggle to.regularwhen Office opens (Dock icon + ⌘-Tab); back to.accessoryon close. Also handles SwiftUI session-restored Office windows via anonAppearhook. - "Office" removed from the mode picker inside the Office window — it was the marker for the window path, not a prompt you'd pick there.
- Office Mode — 7th mode: interactive selection-rewriter. Grab the selection in any app, open a window with the text pre-filled, pick a mode, optionally tweak the input, and paste the result back into the source app on ⌘↵. Clipboard fallback when no selection is active, clipboard auto-copy once processing succeeds, file-drop (txt/md/json/csv/code, 200 KB hard cap) as an alternative input for when you'd rather drop a file than highlight text. Preserves the fire-and-forget
⌘⌥0flow untouched. - Inline connection recovery: when a voice-mode LLM call fails with a recoverable error (connection, auth, 5xx), the HUD no longer hides and loses the transcript. Instead it stays open and shows a profile picker inline — the failed profile is marked and disabled, a 30 s countdown pill ticks down, pick any other profile and Erneut senden retries the same transcript without re-transcribing. The transcript is mirrored to the system pasteboard the instant recovery starts (safety net), and stays there even if the countdown expires.
- Structured LLM errors (
LLMError.swift): every client (Anthropic, OpenAI, Ollama) now throws.connectionFailed/.authFailed/.serverError/.otherinstead of a plainNSError.isRecoverabledecides whether the HUD offers recovery or falls through to the old auto-hiding.fehlerstate. - Non-voice modes are filtered out of voice UI paths: the HUD mode pills and the hotkey dispatcher only iterate
Mode.voiceModes(modes 1–6). Office Mode appears in the menu-bar popover with an Öffnen button instead of Starte, and in Settings → Hotkeys as a normal recorder row.
- Connection profiles — replace the single-provider setup with a full profile system. Each profile holds a provider (Anthropic / OpenAI / Ollama), a base URL, an auth scheme (
x-api-key/Bearer/ none), and a model. Any number of profiles, quick-switch with one click. - New Settings tab: Profile — list, inline editor, JSON import/export, quick-switcher chips, per-profile model discovery (live list from the endpoint).
- Profile scanner — Auf diesem Mac suchen reads
~/.claude-profiles/*.json,~/.claude/settings.json, and~/.config/claude/*.jsonand offers to import them automatically. - LLMRouter — central routing layer that sends every LLM call through the active profile (or falls back to legacy settings for existing installs).
- Keychain — truly silent: all keychain items use an open-access ACL (
SecAccessCreatewith empty trusted-apps list). macOS never prompts for a password or "Allow / Always Allow" confirmation — not on first launch, not after rebuilds, not ever. A one-time migration at first launch rewrites any legacy items; subsequent launches skip it entirely. - Settings window is now resizable — drag to any size; minimum is 780 × 580 px.
- General tab simplified — provider picker and per-provider key fields moved to the new Profile tab.
- Text rewriting via hotkey — no voice, no Services menu required. Select text in any app that supports Accessibility → press the configured hotkey (default
⌘⌥0) → selection gets rewritten using the configured default mode. Result is pasted back over the selection. - Why this replaces v1.0.9's Services approach: macOS Gatekeeper consistently refuses to surface Services from non-notarized apps (
spctl -areturnsrejected, LaunchServices sets thelaunch-disabledflag). The self-signed dev workflow can't overcome that without an Apple Developer account ($99/yr). The hotkey path uses only Accessibility, which is already granted. - Uses AX API first, ⌘C fallback: reads the focused element's
AXSelectedTextattribute directly (works in most native apps). For apps that don't expose AX text (some Electron apps), it simulates ⌘C, reads the clipboard, and restores the previous clipboard contents afterwards — so your clipboard history stays intact. - Setting moved: Settings → Allgemein → „Text umschreiben (Hotkey, ohne Stimme)" — hotkey recorder + default-mode picker (Business/Plus/Rage/Emoji/Prompt).
- NSServices entries and
ServiceProvider.swiftfully removed. The Info.plist is clean. No cross-app registration state remains.
- macOS Services integration — rewrite text without voice. Select text in any app (Mail, Notes, Safari, Pages, most web inputs) → right-click → Services submenu → pick a blitzbot mode. The selection is replaced in place with the rewritten text. Works alongside the existing voice flow — no mic needed.
- Six service entries: blitzbot: Business / Plus / Rage / Emoji / Prompt plus blitzbot: Umschreiben (Default) which uses a configurable default mode.
- New setting in Settings → Allgemein → Kontextmenü: pick the default mode used by the "Umschreiben (Default)" entry + toggle whether to keep the original text in the clipboard on error.
- Language auto-detection: the selected text is routed to the DE or EN system prompt based on a lightweight stop-word analysis, no manual switching needed.
- If a service entry doesn't appear after install: Settings → Allgemein → Kontextmenü → Dienste-Menü neu laden, or log out / back in once. Apple's
pbsdaemon sometimes needs a kick.
- Note on compatibility: works in any app that offers Services on its context menu (TextEdit, Mail, Notes, Pages, Safari text fields, most native apps). Some Electron apps (Slack desktop, Teams) don't surface Services — use the dictation flow there instead.
- More visible waveform amplitude: the yellow HUD waveform now renders peaks ~4.5× taller than before so speech is visually obvious at a glance. Samples are soft-clamped to ±1 so the wave still stays inside the 72 px HUD slot — no visual breakout.
- Idle (post-stop) grey trace also slightly stronger for better continuity between recording and idle states.
- Multi-LLM provider support: switch between Anthropic Claude, OpenAI ChatGPT, and Ollama (local LLM). Per-provider model picker and API key. Normal mode remains local regardless of provider.
- Anthropic: existing flow (Sonnet/Opus/Haiku,
x-api-key). - OpenAI:
gpt-4o-mini/gpt-4o+ free-text model field, bearer auth,/v1/chat/completions. - Ollama: base URL (default
http://localhost:11434), dynamic model list from/api/tagswith green/red health indicator + refresh button, optional bearer auth, 300 s chat timeout for local-model latency.
- Anthropic: existing flow (Sonnet/Opus/Haiku,
- Provider-aware error suppression: the stale "Ollama nicht erreichbar" banner no longer leaks into the menu-bar header when another provider is active. Errors get auto-cleared when the provider changes.
- Provider-aware key warning: menu-bar footer warns about missing Anthropic/OpenAI keys only for the currently selected provider. Ollama shows no key warning (local use often needs none).
- Settings → General restructured: LLM-Provider picker at top of the tab, provider-specific section below (keys, models, URL, health), rest unchanged.
- Real audio waveform: HUD waveform now draws actual PCM sample data from the microphone (Canvas-based, scrolling oscilloscope). Was a simulated bar-chart animation before.
- Waveform always yellow during recording: the waveform line stays consistently yellow for the entire recording session — no more grey flicker during pauses in speech.
- "Stimme erkannt" badge: a small green-dot pill fades in over the waveform as soon as voice input is detected, fades out on silence. Gives immediate confirmation that the mic is picking up audio.
- Pause / Resume: new Pause button in the HUD controls row. Audio engine pauses (keeping the WAV file open), resumes seamlessly. Auto-stop timer resets on resume.
- Auto-Stop countdown clock: circular draining-ring indicator next to the Pause button shows seconds remaining until auto-stop fires. Appears only when the countdown is active.
- "Stille erkannt" delay: the silence banner now waits 5 seconds of continuous silence before appearing (was immediate). Prevents flicker on natural pauses mid-sentence.
- Auto-stop default: 60 seconds (was 45 s). Settings picker now starts at 10 s.
- Cancel button (X): HUD top-left now has an X button to abort a recording without pasting. Previous versions had no discard path.
- No layout jumps: the silence banner area has a reserved fixed height; it fades in/out via opacity (
.easeInOut 0.35 s) instead of adding/removing from layout. - Auto-Execute toggle (↵): per-recording toggle in the HUD bottom row. When on, blitzbot simulates Return after the paste — useful for submitting messages in ChatGPT, Slack, etc. Resets to off on each new recording.
- Security hardening: transcript content no longer written to log file (length only). API error messages sanitized — internal error types don't leak to logs or UI. Anthropic API request timeout set to 120 s.
- Custom prompt append mode: Prompts tab now offers "Replace default" or "Append to default" per mode. Append is useful for "use Business style, but also be informal with colleagues".
- Release build pipeline clarified:
./build-app.sh --releasenow explicitly ad-hoc signs and packages a.zipfor GitHub Releases. The local dev cert (blitzbot-dev) never ends up in a release artifact — it stays on the developer's machine for in-place rebuilds. - README: new First launch on macOS section with the Gatekeeper workaround (right-click → Open, or
xattr -dr com.apple.quarantinefor macOS 15+ "damaged" message), and an honest note about why the app isn't notarized. - Signing identity table added to the Development section so it's clear what each build mode ships.
- No code changes — this is a distribution-hygiene release.
- English input now actually gets English output. Regression fix for the language-routing logic introduced in v1.0.2.
- Root cause:
AppConfig.init()used to eagerly populateprompts[mode]with the German default (via a back-compat accessor that ignored the language parameter).prompt(for:language:)then always took the "custom prompt" branch because the dictionary was non-empty, and the language-aware default on the next line was never reached. The bug was silently re-persisted into UserDefaults on everysave(). - Fix: split custom user overrides (
customPrompts) from defaults. A missing key now means "use language-appropriate default". One-time migration ininit()strips any previously auto-persisted German-default strings so existing installs recover automatically. Real user customizations (differ from the German default) are preserved.
- Root cause:
- Prompts tab in Settings now shows which language's default is active when no override is set, and offers a "Reset to default" button if you did override.
- Mode 6 repurposed: was AI Command (Claude executes the instruction). Now Prompt — Claude turns a loose spoken idea into a clean, tool-agnostic prompt you paste into your AI of choice. Output is the prompt, not the result. Helpful when you know what you want but haven't yet structured it.
- Language detection hardened:
whisper-cli -l automislabels short English utterances as German more often than expected. v1.0.3 adds a stop-word-ratio content detector on top: if the transcript clearly looks English or German, that overrides Whisper's metadata. Prompt routing now follows the content, not Whisper's self-assessment.
- AI Command mode (6th mode,
⌘⌥6): dictate an instruction and Claude executes it instead of polishing the wording. Write code, run an analysis, answer a factual question — the result gets pasted directly. Ideal for quick "write me a Python function that…" or "summarize this in three bullets" without opening a chat UI. - Automatic language detection: Whisper now auto-detects the spoken language (previously hardcoded to German). The Claude polish runs in the same language. German and English prompts ship as defaults for all modes.
- Manual language override: Settings → General → Output language with Auto / Deutsch / English — forces transcription and Claude output to a specific language regardless of what was spoken.
- HUD language badge: while recording, a small
DEorENpill shows next to the mode name so you see what will actually be written. - HUD width grew from 480 → 560 px to fit the 6th mode pill cleanly.
- HUD mode switcher: five clickable mode pills + a red Stop button inside the floating recording HUD. Switch modes or end the recording without reaching for the keyboard. Panel still doesn't steal focus.
- Cmd+Q / Quit fixed: the Beenden button in the menu bar popover now properly terminates the app under the accessory activation policy (
keyboardShortcut("q", modifiers: .command)added). - Settings UI refactor: replaced the stock macOS
TabViewwith a custom icon toolbar at the top of the settings window. Six colored tiles (General/Hotkeys/Prompts/Vocabulary/Setup/About), clearer affordance, more native look. - Hotkey migration: when adding Business at position 2, the existing stored bindings for Plus/Rage/Emoji would have collided with the new Business default (⌘⌥2) and left Emoji unbound. v1.0.1 resets all mode shortcuts to the v1.0.1 defaults exactly once on first launch, guarded by a one-time migration flag. Subsequent user customizations persist normally.
- Build hygiene:
./build-app.shnow builds into~/Downloads/blitzbot-build/viaswift build --scratch-pathinstead of leaving.build/inside the (possibly sync'd) project directory. Also updates/Applications/blitzbot.appin place.
Initial public release.
- 5 modes: Normal (offline), Business, Plus, Rage, Emoji with individual hotkeys (default
⌘⌥1-5) - Live mode switch while recording
- Floating HUD with timer + live waveform + mode badge
- Custom hotkeys per mode via KeyboardShortcuts recorder
- Editable system prompts per mode
- Vocabulary list fed to Whisper as context for proper nouns and jargon
- Auto-updater via GitHub Releases (check from Settings → About)
- Setup wizard for macOS permissions
- DE + EN UI
- Persistent log at
~/.blitzbot/logs/blitzbot.log - MIT license
MIT — see LICENSE.
© 2026 blitzbot contributors.