Marvis v0.1.4

What's Changed

• chore: relicense from MIT to Apache-2.0 by @chenillen in #41
• Prompt Presets for Chat by @chenillen in #42
• feat: transcript export and audio playback with export file by @chenillen in #44
• feat: rename diarized speakers and with image attachments by @chenillen in #45
• full refactor by @chenillen in #46
• chore: upgrade @opennextjs/cloudflare to 1.20.9 and patch preview-props manifest loading by @chenillen in #47
• docs: mark Phase 1 complete and reorganize roadmap structure by @allengineerx in #48
• feat: add local memory extraction, storage, and profile support by @chenillen in #49
• session compaction for short turn memory, optimize replay of audio session by @chenillen in #50
• fix: parse codesign entitlements in both text and XML formats by @chenillen in #51

New Contributors

• @allengineerx made their first contribution in #48

Full Changelog: v0.1.2...v0.1.4

Automated multi-platform build of Marvis — the Tauri app with the bundled whisper-cli sidecar.

Downloads

Platform	File	Notes
macOS Apple Silicon	*_aarch64.dmg	Ad-hoc signed — right-click → Open (or xattr -cr Marvis.app)
macOS Intel	*_x64.dmg	Ad-hoc signed — right-click → Open
Windows x64	*-setup.exe	NSIS installer, unsigned — SmartScreen will warn
Windows x64	*_en-US.msi	MSI installer, unsigned
Linux x64	*.AppImage	chmod +x and run
Linux x64	*.deb	sudo dpkg -i

SHA256SUMS.txt covers all bundle files.

Caveats

• Unsigned test builds — no Developer ID signing, notarization, or Gatekeeper pass.
• Packages built by marvis-build.yml package jobs for tag v0.1.4.

Marvis v0.1.2

Automated multi-platform build of Marvis — the Tauri app with the bundled whisper-cli sidecar.

Downloads

Platform	File	Notes
macOS Apple Silicon	*_aarch64.dmg	Ad-hoc signed — right-click → Open (or xattr -cr Marvis.app)
macOS Intel	*_x64.dmg	Ad-hoc signed — right-click → Open
Windows x64	*-setup.exe	NSIS installer, unsigned — SmartScreen will warn
Windows x64	*_en-US.msi	MSI installer, unsigned
Linux x64	*.AppImage	chmod +x and run
Linux x64	*.deb	sudo dpkg -i

SHA256SUMS.txt covers all bundle files.

Caveats

• Unsigned test builds — no Developer ID signing, notarization, or Gatekeeper pass.
• Packages built by marvis-build.yml package jobs for tag v0.1.2.

Marvis v0.1.1

Automated multi-platform build of Marvis — the Tauri app with the bundled whisper-cli sidecar.

Downloads

Platform	File	Notes
macOS Apple Silicon	*_aarch64.dmg	Ad-hoc signed — right-click → Open (or xattr -cr Marvis.app)
macOS Intel	*_x64.dmg	Ad-hoc signed — right-click → Open
Windows x64	*-setup.exe	NSIS installer, unsigned — SmartScreen will warn
Windows x64	*_en-US.msi	MSI installer, unsigned
Linux x64	*.AppImage	chmod +x and run
Linux x64	*.deb	sudo dpkg -i

SHA256SUMS.txt covers all bundle files.

Caveats

• Unsigned test builds — no Developer ID signing, notarization, or Gatekeeper pass.
• Packages built by marvis-build.yml package jobs for tag v0.1.1.

v0.1.0

v0.1.0

The first public release of Marvis — a privacy-first AI co-pilot that
lives on your desktop. Marvis floats above your workspace as a small
translucent bar: it can see your screen (with your permission), hear your
meetings, and stream answers into an overlay — while your API keys,
history, and screen data stay entirely on your machine. No accounts, no
cloud sync; nothing leaves your device except the requests you send to the
providers you choose.

Highlights

The bar

• A frameless, always-on-top overlay that morphs from a breathing capsule
  into an input pill, then into a chat/listen card.
• Docks to any screen edge, drags anywhere, remembers its position, and
  can be locked in place.
• Right-click context menu and tray icon with live state — start/stop
  capture, start a Listen, open history, position, and lock.
• Liquid-glass material on macOS; launch wordmark intro; lazy-loaded views
  keep the webview light.

Ask — screen-aware questions

• Type a question and press Cmd/Ctrl+Enter to send it with the latest
  screen frame; answers stream back as markdown.
• Screen content lives in an in-memory ring buffer (~60 s horizon, capped
  at 120 frames / 64 MB) — screenshots are never written to disk.
• Choose exactly what to share: a native content-sharing picker offers
  displays, windows, and thumbnails before capture starts.
• An optional vision provider can read the frame first and answer over its
  text description.
• Dictate straight into the input with the push-to-talk mic.
• Token usage is tracked per message, with a retry command for failed
  sends.

Listen — meeting transcription

• Dual-channel capture: your mic ("me") + system audio ("them") flow into
  streaming speech-to-text.
• Live diarized transcript with speaker labels, color-coded turns, speaker
  filters, elapsed timer, pause/resume, and one-click transcript copy.
• Rolling AI summaries pin a TLDR strip above the transcript, refreshed
  every five turns — with follow-up chips that ask questions against the
  meeting's own chat context.
• Optional voice enrollment identifies you as "You" across sessions and
  languages.
• Each session can record a single 16 kHz mono WAV mixing both channels.

Bring your own model

• LLM providers: OpenAI, Anthropic, Gemini, OpenRouter,
  Ollama (no API key needed), or any
  OpenAI-compatible endpoint.
• Providers form a drag-to-reorder failover chain — a failed call hands
  off to the next configured provider.
• Speech-to-text: Deepgram streaming (interim results), fully-local
  bundled whisper.cpp (tiny / base / small models downloaded on demand
  from Hugging Face), or sherpa-onnx SenseVoice with an optional English
  punctuation model and built-in diarization.
• A main-language preference drives chat, summary, and transcription
  output.

History & settings

• Sessions, messages, transcripts, and summaries persist locally — reopen
  any chat or listen session where you left off.
• Guided first-run onboarding: appearance → screen consent → provider key
  (validated before saving, or pick an Ollama model) → optional local
  voice models.
• A liquid-glass settings window covers appearance, providers, recording,
  hotkeys, and language.
• marvis:// deep links — marvis://ask?text=... focuses the bar and
  fires an Ask; any other link surfaces the app.

Hotkeys

Five rebindable global chords (Settings → Hotkeys or [hotkeys] in
~/.marvis/config.toml):

Action	Default
Show / hide the input	Cmd/Ctrl+Alt+Space
Toggle screen capture	Cmd/Ctrl+Alt+R
Start a Listen	Cmd/Ctrl+Alt+T
Open history	Cmd/Ctrl+Alt+H
Lock bar position	Cmd/Ctrl+Shift+L

Fixed in-bar keys: Enter send · Shift+Enter new line ·
Cmd/Ctrl+Enter send with screenshot · Cmd/Ctrl+, settings.

Platform support

Platform	Screen capture	System audio	Consent
macOS (Apple Silicon / Intel)	ScreenCaptureKit	ScreenCaptureKit	Screen Recording permission
Windows (x86_64)	Windows.Graphics.Capture	WASAPI loopback	none — WGC is the consent boundary
Linux (x86_64)	XDG portal + PipeWire	PulseAudio monitor	portal picker (persisted token)

Downloads

Platform	Artifacts
macOS Apple Silicon	Marvis_0.1.0_aarch64.dmg
macOS Intel	Marvis_0.1.0_x64.dmg
Windows x64	Marvis_0.1.0_x64-setup.exe (NSIS), Marvis_0.1.0_x64_en-US.msi
Linux x64	Marvis_0.1.0_amd64.AppImage, Marvis_0.1.0_amd64.deb

Privacy notes

Everything Marvis writes lives under ~/.marvis
(%USERPROFILE%\.marvis on Windows), created 0700 where the OS has
mode bits:

• keys.json (0600) — plaintext provider API keys inside the 0700
  root; there is no encryption in this build. Keys are never serialized
  to the UI (masked …last4 only) and never logged.
• config.toml — non-secret preferences.
• marvis.db (0600) — SQLite sessions, messages, transcripts,
  summaries.
• models/ — downloaded whisper / sherpa voice models.

Security issues: please report privately to support@getmarvis.com.

What's next

Session history UI improvements, prompt presets, Ollama model management,
and a Gemini search-grounding toggle are on the
roadmap. Bug reports and
feature requests →
GitHub issues.