Documentation

NimbleScribe Docs

Everything you need to install, configure, and get the most out of NimbleScribe — free offline voice transcription and local AI for macOS.

Quick Start Voice Dictation Meetings Nimble AI Troubleshooting

Overview

What is NimbleScribe?

NimbleScribe is a free, fully offline voice productivity app for macOS. It lets you dictate text directly into any application by pressing a keyboard shortcut, record meetings silently, and chat with an on-device AI assistant — all without sending a single byte of data to any server.

NimbleScribe is built on a model-agnostic architecture. The transcription engine currently supports the Whisper and Parakeet families of speech recognition models, with more model integrations — including cloud-optional providers — planned for future releases. The AI assistant similarly runs on local, swappable language models.

Transcription models

Whisper · Parakeet · more

AI assistant

On-device LLM

Languages supported

100+ (Whisper) / EN (Parakeet)

Network calls after setup

Requirements

System requirements

macOS13 Ventura or later (14 Sonoma, 15 Sequoia fully supported)ChipApple Silicon — M1, M2, M3, or M4 (all variants: Pro, Max, Ultra)RAM8 GB minimum · 16 GB recommended when running large models + AI simultaneouslyDisk space~200 MB for the app · 300 MB – 2.2 GB per transcription model · ~500 MB for AI modelInternetRequired only for the initial model download — permanently offline after thatArchitectureIntel Macs are not supported — performance requires the Apple Neural Engine

Intel Macs:NimbleScribe uses Apple's Core ML and the Neural Engine for real-time on-device inference. These are not available on Intel processors, which is why only Apple Silicon is supported.

Windows

WindowsWindows 10 or Windows 11 (x64)ChipAny modern x64 CPU — Intel or AMDRAM8 GB minimum · 16 GB recommended when running large models + AI simultaneouslyDisk space~200 MB for the app · 300 MB – 2.2 GB per transcription model · ~500 MB for AI modelInternetRequired only for the initial model download — permanently offline after thatInstallMSIX package — enable Developer Mode or Sideload apps in Windows Settings before installing

Installation

Installation & setup

1

Download NimbleScribe

Download NimbleScribe.dmg from the link above — it's free, no email or account required.

2

Install the app

Open the .dmg file. Drag NimbleScribe.app to your Applications folder. On first launch, macOS may show a security warning — click Open to proceed. This is expected for apps distributed outside the App Store.

3

Grant permissions

NimbleScribe will prompt for three permissions on first launch:

Microphone — required for voice dictation.
Accessibility — required to paste text at your cursor position in any app.
Screen Recording — required for the meeting recorder to capture system audio.

4

Download a transcription model

Open Settings → Models. Choose a model family (Whisper or Parakeet) and select a size. Click Download — this is a one-time download that stays on your Mac permanently. We recommend Whisper Large-v3-turbo for multilingual use, or Parakeet TDT 1.1B for English-only high-speed dictation.

5

Start transcribing

Click inside any text field in any app. Press your shortcut (default: ⌃⇧ Space), speak clearly, then press it again to stop. Your transcription appears at the cursor instantly.

First Use

Your first transcription

Once a model is downloaded, using NimbleScribe is intentionally simple:

Click to focus a text field in any app (Slack, Notion, email, terminal — anything).

Press your recording shortcut and begin speaking. A recording indicator appears in the menu bar.

Speak naturally at a normal pace. The transcription engine handles punctuation and context automatically.

Release the shortcut or press it again to stop. Your text appears at the cursor within 500 ms.

The first transcription after app launch may take 1–2 seconds longer as the model loads into the Neural Engine. Subsequent transcriptions are instant.

Core Features

Voice Dictation

NimbleScribe's core feature. Press your shortcut anywhere in macOS, speak, release — text appears at your cursor. It uses the macOS Accessibility API for direct injection, meaning text goes straight to the active text field without touching the clipboard.

How it works

✓Your microphone audio is captured in real time.

✓The selected transcription model (running locally) converts speech to text.

✓The transcribed text is injected directly at the keyboard focus point using the macOS Accessibility API.

✓No clipboard is modified. Undo (⌘Z) works as expected.

Compatible apps

NimbleScribe works in any macOS app that accepts keyboard input, including:

SlackNotionVS CodeXcodeTerminalSafariChromeMailNotesMicrosoft WordGoogle DocsLinearFigmaObsidianBearRaycastand everything else

Note: If text does not paste in a specific app, make sure NimbleScribe has Accessibility permission enabled in System Settings → Privacy & Security → Accessibility.

Shortcuts

Keyboard Shortcuts

All recording shortcuts are fully customisable. NimbleScribe registers them as system-wide global hotkeys that fire even when the app is in the background, the window is hidden, or you are in full-screen mode.

Shortcut	Action	Notes
`⌃⇧ Space`	Start / stop voice dictation	Default — fully configurable
`Custom combo`	Meeting record start / stop	Set in Settings → Shortcuts
`⌘ ,`	Open Settings
`⌘ K`	Open Command Palette / Search	From main app window
`Click menu icon`	Open menu bar popover	Quick access without opening window

To change a shortcut, open Settings → Shortcuts, click in the shortcut field, and press your desired key combination. NimbleScribe detects it immediately.

Models

Transcription Models

NimbleScribe is model-agnostic. It currently supports two families of speech recognition models, with more integrations planned. All models run entirely offline after a one-time download. You can switch models at any time in Settings → Models.

Whisper (OpenAI) supports 100+ languages and is best for multilingual use. Parakeet (NVIDIA) is English-only but significantly faster and more accurate for English dictation.

Family	Model	Size	Speed	Accuracy	Languages
Whisper	Tiny	~300 MB	Fastest	Good	100+
Whisper	Base	~400 MB	Very fast	Better	100+
Whisper	Small	~500 MB	Fast	Great	100+
Whisper	Medium	~750 MB	Moderate	Excellent	100+
Whisper	Large-v3-turborecommended	~1.5 GB	Fast (turbo)	Best	100+
Parakeet	TDT 0.6B	~1.2 GB	Extremely fast	Excellent	English
Parakeet	TDT 1.1B	~2.2 GB	Very fast	Best-in-class	English

More model families — including additional multilingual models and optional cloud-connected providers — are planned for future releases. You can delete unused models from Settings → Models at any time to reclaim disk space.

Meeting Recorder

NimbleScribe can record your meetings silently — capturing full system audio from Zoom, Teams, Google Meet, or any other platform. There is no bot, no shared link, and no notification to other participants. The recording stays entirely on your Mac.

How it works

✓NimbleScribe detects active Zoom, Teams, or Google Meet windows automatically.

✓It shows a recording prompt — click to start. Or use your configured meeting shortcut.

✓System audio (your mic + everyone else on the call) is captured using macOS Screen Recording permission.

✓When you stop, the selected transcription model processes the full recording with speaker-separated timestamps.

✓Nimble AI then generates a structured summary, key decisions, and action items automatically.

What you get after a meeting

Full transcript

Every word, timestamped

Meeting summary

2–3 paragraph AI overview

Key decisions

Extracted from discussion

Action items

Who, what, by when

Privacy:Meeting audio is captured and stored locally on your Mac only. No audio, transcript, or summary is ever transmitted to any server. Recordings are stored in your Mac's Application Support folder and can be deleted from the Meetings tab.

Nimble AI

Nimble AI — Local AI Assistant

Nimble AI is an on-device AI assistant that can answer questions about your transcriptions and meetings in plain language. It runs a local, quantised language model entirely on your Mac — no API key, no cloud account, no internet connection required.

Like the transcription engine, the AI assistant is model-agnostic. The current default model is a compact, high-quality LLM optimised for Apple Silicon. Additional model options and future cloud-optional providers are planned.

What you can ask

"What were the action items from the meeting?"

"Summarise the last 3 transcriptions in bullet points."

"Draft a follow-up email based on the discussion."

"What decisions did we make about the Q3 roadmap?"

"Translate this transcription into English."

"What did I say about the budget in my last meeting?"

Setup

1

Download the AI model

Open Settings → AI Engine and click Download AI Model. This is a one-time download.

2

Open the Nimble AI sidebar

Click the Nimble AI button in the toolbar.

3

Start a conversation

Type your question. Nimble AI has access to all your transcriptions and meeting notes — ask in plain language.

Performance: Nimble AI works best with 16 GB RAM. On 8 GB systems, responses may be slightly slower. The model runs on the Neural Engine for minimal CPU and battery impact.

AI Cleanup

Grammar AI Cleanup

After transcription, Grammar AI Cleanup passes the raw text through a second on-device language model. It removes filler words ("um", "uh", "like"), corrects punctuation, fixes capitalisation, and smooths repetitions — without changing your actual words or meaning. Completely offline.

Before / After

Raw transcription

"um so basically we need to uh finish the the design system by friday and uh yeah schedule those user interviews"

After cleanup

"We need to finish the design system by Friday and schedule those user interviews."

How to enable

Open Settings → AI Engine → Grammar Cleanup. Download the cleanup model (one-time, ~500 MB), then toggle Enable Grammar Cleanup on.

Grammar Cleanup adds a short post-processing step. If you dictate code or technical strings that should not be altered, disable Cleanup temporarily in Settings.

History & Export

Transcription History

Every dictation and meeting is saved automatically to a local SQLite database on your Mac. Access your full history from the Transcriptions or Meetings tab.

Search

Use ⌘ K or the search icon to search across your entire history by keyword. Full-text, instant, entirely on-device.

Export

Select one or multiple transcriptions. Use the export button to copy as plain text or download as a .txt file. Meeting transcripts include timestamps and AI-generated notes.

Auto-delete policy

Configure cleanup in Settings → General → History. Options: 1 month, 3 months, 6 months, 1 year, or never. Audio recordings can be retained or deleted separately from text transcripts.

Productivity

Menu Bar & Popover

NimbleScribe lives permanently in the macOS menu bar, even when the main window is closed. Click the icon to open the quick popover — a compact panel that gives you immediate access without switching context.

✓View your 5 most recent transcriptions with one-click copy.

✓Full-text search across your entire transcription history.

✓Quick Settings shortcut to open the main settings window.

✓Recording status indicator — shows when NimbleScribe is actively listening.

✓Option to hide the app from the Dock entirely (menu bar only mode).

Go to Settings → General → Show in Dock to control whether NimbleScribe appears in the Dock. When hidden, the menu bar icon is your only entry point.

Vocabulary

Custom Vocabulary

For technical terms, product names, or domain-specific vocabulary that the transcription engine might not recognise correctly, you can add custom words to improve accuracy.

1

Open Vocabulary settings

Go to Settings → AI Engine → Vocabulary.

2

Add custom words

Click Add word and type the term exactly as you want it to appear. You can add company names, product names, acronyms, or technical jargon.

3

Verify accuracy

Dictate a phrase containing the word and verify the output. If accuracy is still low, try adding phonetic alternatives or common misspellings.

Appearance

Themes & Appearance

NimbleScribe ships with four visual themes. Switch at any time from Settings → General → Theme.

Dark

Deep dark interface. Best for low-light environments and extended use.

Light

Clean light interface. Best for bright environments or lighter macOS setups.

System

Follows your macOS appearance preference automatically.

Glass

Frosted glass translucency. Adapts to your desktop wallpaper.

Privacy & Security

NimbleScribe's privacy guarantee is architectural, not policy-based. There is no server to send data to — all processing happens locally.

✓Audio never leaves your Mac — ever.

✓Transcriptions stored in a local SQLite database, not synced anywhere.

✓No account, no login, no profile.

✓No usage analytics, no telemetry, no crash reports.

✓No third-party SDKs that could exfiltrate data.

✓Only two network calls in the app's lifetime: update check (dismissable) and model download (manual).

✓Fully operable in air-gapped environments after model download.

Regulated industry use

Because NimbleScribe has no data egress, it requires no GDPR data processing agreements, no HIPAA BAA, and no security review process. There is no data to disclose. NimbleScribe is safe to use for healthcare notes, legal transcription, financial dictation, or any other sensitive context.

You can verify NimbleScribe's network activity using macOS's built-in Network Monitor in Activity Monitor or a tool like Little Snitch. You will see zero outbound calls during transcription.

Permissions

macOS Permissions

NimbleScribe requests three macOS permissions. Here is exactly why each is needed:

Microphone

System Settings → Privacy & Security → Microphone

Required to record your voice for dictation and meeting capture.

Accessibility

System Settings → Privacy & Security → Accessibility

Required to inject transcribed text directly at your cursor position in any application. Without this, NimbleScribe cannot paste into other apps.

Screen Recording

System Settings → Privacy & Security → Screen Recording

Required for the Meeting Recorder to capture system audio. NimbleScribe does not record your screen — only the audio stream.

Troubleshooting

Common Issues

Shortcut does not work in some apps

NimbleScribe needs Accessibility permission to paste into other applications. Go to System Settings → Privacy & Security → Accessibility and ensure NimbleScribe is toggled on.

Microphone not detected or no audio captured

Check System Settings → Privacy & Security → Microphone. Ensure NimbleScribe is permitted. Also verify the correct input device is selected in System Settings → Sound.

Meeting recorder captures no audio

The meeting recorder requires Screen Recording permission to access system audio. Go to System Settings → Privacy & Security → Screen Recording and enable NimbleScribe. A restart may be required.

Model download stuck or failed

Ensure you have enough free disk space (check model sizes in the Transcription Models section) and a stable internet connection. Retry from Settings → Models. If the issue persists, delete the partial download and try again.

Transcription is slow or takes several seconds

Smaller models transcribe faster. Try switching to Tiny or Base in Settings → Models. The very first transcription after app launch may take slightly longer as the model loads into the Neural Engine.

Text pasted in the wrong location

Click to focus the exact text field before pressing your shortcut. NimbleScribe pastes at whatever element holds keyboard focus when you stop recording. Make sure you haven't accidentally clicked elsewhere.

Nimble AI gives slow or no responses

The AI model needs to load into memory on first use. Wait 5–10 seconds for the first response. If it never responds, go to Settings → AI Engine and confirm the model is downloaded. Ensure you have at least 8 GB RAM free.

App shows "No model downloaded" warning

Open Settings → Models and download at least one transcription model. We recommend Whisper Large-v3-turbo for the best multilingual experience, or Parakeet TDT 1.1B for English-only use. If you have already downloaded a model, try clicking the refresh button in the Models tab.

Support

Getting Help

NimbleScribe is maintained by the team at Nimble Software Lab. The best way to get help, report a bug, or request a feature is via GitHub Issues.

Report a Bug

Open a GitHub Issue with steps to reproduce. We respond within 48 hours.

Request a Feature

Have an idea? Open a feature request issue on GitHub. We read every one.

Made by Nimble Software Lab

Enterprise software studio based in Dhaka, Bangladesh.
50+ products shipped · 12 countries served · 98% client retention.

Visit NimbleSL.com Download NimbleScribe Free