Settings reference
Every tab in the Fluister settings window, top to bottom.
Fluister's settings window has eight tabs. Open it from the menu bar icon → Settings (or ⌘,).
General
- Appearance — System, Light, or Dark. Match the system, or pick one explicitly.
- Overlay position — pick which corner of the screen the recording pill appears in (3×2 grid).
Recording
Fluister uses your system's default input device. Change it in macOS: System Settings → Sound → Input.
- Spoken language — primary language for transcription. Pick the closest match for best Whisper accuracy. The full multilingual model supports 99 languages.
- Auto-stop after silence — optional. When on, Fluister ends the recording if you go quiet for the chosen window (1.0s, 1.5s, 2.0s, or 3.0s). Off by default; release the hotkey to stop manually.
Cleanup
- AI cleanup — toggle the LLM pass on or off. Removes fillers, fixes punctuation, applies the active profile's style.
- Backend — Bundled (the in-app
llama-serverwith the bundled model) or Ollama (an external Ollama instance you already run on your Mac). Both run locally. - Bundled model — currently Llama 3.2 3B Instruct (Q4_K_M, 2.0 GB). Re-download if the file is corrupted or missing.
Profiles and vocabulary are edited as markdown files in the vault. See Vault and Cleanup → Profiles.
Hotkeys
- Dictation — Right Option. Hold to talk, release to paste. Tap Escape while held to cancel.
- Cancel — Escape (while the dictation key is held). Discards the recording without pasting.
- Customisation — coming soon. Rebinding isn't available yet; Right Option is fixed for now.
Models
The speech-to-text models. Smaller is faster and uses less memory; larger transcribes more accurately, especially with accents, names, and technical words.
Each model ships in two flavours:
- English only — leaner, English-only.
- Multilingual — 99 languages.
Sizes (English / Multilingual):
| Size | English | Multilingual |
|---|---|---|
| Tiny | 39 MB | 39 MB |
| Base | 148 MB | 148 MB |
| Small | 488 MB | 488 MB |
| Medium | larger | larger |
| Large-v3 | largest | largest |
Download what you need on demand. Only the active model is loaded into memory.
Storage
Manages the vault (the markdown mirror of your profiles and vocabulary).
- Vault → Change folder… — pick or move the folder Fluister mirrors data into. Profiles and vocabulary are stored as plain markdown files there. Sync via iCloud, Dropbox, or Git; the data stays on your machine.
- Folder — the current vault path.
- Profiles / Vocabulary — counts of what's currently mirrored.
- Status — Reachable / Unreachable indicator for the vault folder.
- In Finder → Open in Finder — open the vault folder to edit files directly.
- Disable vault — drop back to SQLite-only mode. Files on disk are kept; they just stop being kept in sync.
Integrations
- Enable MCP server — toggle. Lets Claude Desktop, Cursor, Claude Code, and Zed pull your dictation history, run cleanup on arbitrary text, and append notes into the vault. Bound to
127.0.0.1only, never accessible over the network. Off by default. - Endpoint — pick Cursor / Code / Zed or Claude Desktop to show the matching JSON snippet, then Copy snippet into your client's MCP config. Full reference: MCP server.
- What it exposes —
search_dictations,recent_dictations,get_dictation,clean_text,append_to_vault. All local.
About
- App version + update status. The button reads Check for updates…; below it says whether you're up to date.
- Re-run onboarding — runs the first-launch flow again (useful if you denied a permission).
- Visit website — opens the marketing site.
- Source on GitHub — opens the repo.
- License: MIT.
Updating
Fluister checks for updates on launch (silently) and shows a banner if one is available. You can also check manually in Settings → About.
Click Install and restart to download the new version, install it, and relaunch. The update is signed and verified before install.
If the update appears to hang at "Installing — restarting Fluister...", the most likely cause is an open MCP session holding a long-lived stream. Workarounds:
- Toggle MCP off in Settings → Integrations before updating, then back on after.
- Quit your MCP client (Claude Desktop, Claude Code, Cursor) before updating.
This was fixed in v0.1.11. Once you're on that version or later, updates will complete cleanly even with active MCP sessions.