- Model browser: per-million input/output price on every row; time to first token (p50) and tokens/s for the busiest OpenRouter provider, fetched lazily for visible rows and cached 10 minutes. - Fix: Free badge and free-first sorting (prices are strings), and the list was empty until typing. - Settings: text-only tabs styled like Waypoint's; drop the heading that repeated the tab name.
12 KiB
NibbleAI
AI-powered tools for your Obsidian vault — speech-to-text, proofreading, title generation, property filling, and more.
NibbleAI is a central hub for small, practical AI-assisted tasks inside Obsidian. Instead of installing a dozen single-purpose plugins, NibbleAI gives you one consistent interface with shared configuration, a unified AI service, and a growing set of capabilities.
Features
🎙️ Speech-to-Text (Transcription)
Record audio directly in Obsidian or upload existing audio/video files, and get instant transcripts via OpenAI Whisper (or any compatible API — Groq, Azure, etc.).
- Record with your microphone using the built-in recorder
- Upload audio/video files from disk
- Transcribe existing audio files via the file context menu
- Post-process transcripts with an LLM to fix grammar, remove filler words, and format as Markdown
- Auto-generate note filenames from the transcript content
- Template system for note content (
{{date}},{{time}},{{title}},{{transcription}},{{audioFile}})
✍️ Proofreading
Fix punctuation and grammar of selected text with a single command or right-click.
- Prompt injection resistant — the built-in system prompt explicitly guards against prompt injection attacks in the text being proofread
- Smart quote replacement — straight quotes → curly quotes (except in code blocks)
- Timing and token report — a live timer runs while the model works; when done, a notice shows elapsed time, model, upstream host (OpenRouter), and prompt/output/reasoning token counts (also logged to the console)
- Works with any OpenAI-compatible provider (OpenRouter, OpenAI, etc.)
🏷️ Title Generation
Generate a descriptive title for the current note with one command. Choose how to apply it:
- Rename the file — the note's filename becomes the generated title
- Set a frontmatter property — adds or updates a
title:field in the YAML frontmatter
📋 Property Filling
Fill any frontmatter property with AI-generated content. Built-in quick commands for common properties:
- Fill any property — prompts you for the property name
- Fill "summary" — one-click command
- Fill "tags" — one-click command
- Customizable generation prompt with
{property}placeholder
Installation
From Obsidian Community Plugins (once published)
- Open Settings → Community plugins
- Disable Safe mode
- Click Browse and search for "NibbleAI"
- Install and Enable
Manual Installation
- Download the latest release from the Releases page
- Extract the folder to
YourVault/.obsidian/plugins/nibbleai/ - Enable the plugin in Settings → Community plugins
Build from Source
git clone https://github.com/your-username/nibbleai.git
cd nibbleai
npm install
npm run build
Then copy the main.js, manifest.json, and styles.css to your vault's plugins folder.
Getting Started
1. Configure API Keys
Open Settings → NibbleAI and, for at least one provider, pick the secret holding its key from Obsidian's keychain (Settings → Keychain):
| Key | Used For |
|---|---|
| Groq API Key | Whisper transcription, and as a chat provider for proofreading/title/property/post-processing |
| OpenAI API Key | Whisper transcription, and as a chat provider for proofreading/title/property/post-processing |
| Anthropic API Key | Claude models for proofreading/title/property/post-processing |
| OpenRouter API Key | Access to hundreds of models for proofreading/title/property/post-processing |
| Custom API Key + URL | Any OpenAI-compatible endpoint (self-hosted, Azure, etc.) |
Every LLM-backed feature (Proofreading, Title Generation, Property Filler, Post-Processing) shows the same provider dropdown — pick any provider that has a key configured above.
Tip: An OpenRouter API key with a free model like
google/gemini-2.5-flash-litecosts nothing and works great across all four AI features.
2. Try the Features
| Feature | How to Use |
|---|---|
| Record audio | Click the 🤖 ribbon icon, or use the Start/stop recording command (Alt+Q) |
| Upload audio | Use the Upload audio file command |
| Transcribe existing file | Right-click an audio file in the file explorer |
| Proofread text | Select text → right-click → "Fix punctuation and grammar with AI" |
| Generate title | Use the Generate title for current note command |
| Fill property | Use the Fill property with AI command (or the presets for summary / tags) |
Commands
| Command | Icon | Default Hotkey | Description |
|---|---|---|---|
| Start/stop recording | mic |
Alt+Q |
Toggle recording on/off |
| Upload audio file | upload |
— | Choose an audio/video file to transcribe |
| Pause/resume recording | pause |
— | Pause or resume an active recording |
| Open recording controls | popup-open |
— | Show the recording modal |
| Fix punctuation and grammar | wand |
— | Proofread selected text with AI |
| Generate title for current note | heading |
— | Generate and apply a title |
| Fill property with AI | pencil |
— | Prompt for a property name, then fill it |
| Fill "summary" property | list |
— | Directly fill the summary property |
| Fill "tags" property | tags |
— | Directly fill the tags property |
URI Protocol
NibbleAI registers the nibbleai:// protocol handler, allowing external tools to control recording:
| URI | Action |
|---|---|
obsidian://nibbleai |
Open recording controls |
obsidian://nibbleai?command=start |
Start recording |
obsidian://nibbleai?command=stop |
Stop recording |
obsidian://nibbleai?command=pause |
Pause/resume recording |
obsidian://nibbleai?command=cancel |
Cancel recording |
This is useful for automation tools, Stream Deck, or scripting.
Model Selection
Wherever you need to select an AI model (post-processing, proofreading, title generation, property filler), NibbleAI provides a Browse button that opens a searchable dialog of available OpenRouter models:
- Models are fetched live from
openrouter.ai/api/v1/models - Free models appear first (marked with a Free badge)
- Each row shows the price per million input/output tokens
- With an OpenRouter key configured, each visible row also shows time to first token (p50) and output speed (tokens/s) over the last 30 minutes, for the provider that served the most requests. Loaded once you stop typing; cached for 10 minutes
- Type to filter by name, ID, or description
- You can still type a model ID manually if you prefer
Settings Reference
API Keys
Keys live in Obsidian's keychain (SecretStorage). NibbleAI stores only the name of the secret you picked for each provider in data.json, and reads the key at the moment it calls the API, so a key rotated in the keychain applies immediately. Upgrading from ≤ 1.2.1 links the previously used groq-api-key, openai-api-key, anthropic-api-key, openrouter-api-key and custom-api-key secrets automatically.
Transcription
| Setting | Default | Description |
|---|---|---|
| API URL | https://api.openai.com/v1/audio/transcriptions |
Whisper-compatible endpoint |
| Model | whisper-1 |
Transcription model |
| Language | (empty) | Language hint (auto-detect if empty) |
| Prompt | (empty) | Boosts accuracy for specific terms |
| Cursor context | Off | Send text around cursor as context |
| Temperature | 0 |
Sampling temperature (0–1) |
| Response format | json |
Output format |
Recording
| Setting | Default | Description |
|---|---|---|
| Microphone | Default | Audio input device |
| Save audio file | On | Save recording to vault |
| Audio save path | (vault root) | Folder for audio files |
Output
| Setting | Default | Description |
|---|---|---|
| Create note file | On | Create a .md file per transcription |
| Note save path | (vault root) | Folder for notes |
| Note filename template | {{datetime}} |
Variables: {{date}}, {{time}}, {{datetime}}, {{title}} |
| Note template | ![[{{audioFile}}]]\n{{transcription}} |
Variables: {{transcription}}, {{audioFile}}, {{title}} |
Post-Processing
| Setting | Default | Description |
|---|---|---|
| Enable | Off | Clean up transcripts with an LLM |
| Provider | Anthropic | Any configured provider: Anthropic, OpenAI, Groq, OpenRouter, or Custom |
| Model | claude-sonnet-4-20250514 |
LLM for cleanup |
| Auto-generate title | Off | Create filenames from content |
Proofreading
| Setting | Default | Description |
|---|---|---|
| Provider | OpenRouter | Any configured provider: Anthropic, OpenAI, Groq, OpenRouter, or Custom |
| Model | openai/gpt-4o-mini |
Model for proofreading |
| Temperature | 0.2 |
Lower = more consistent |
| Max tokens | 2000 |
Maximum response length |
| Reasoning | Off | OpenRouter only. Let the model think before answering; slower and uses more tokens |
| Effort | Low | OpenRouter only. Output effort (verbosity); with reasoning on, also reasoning depth |
| System prompt | (built-in) | Editable with prompt injection protections |
Title Generation
| Setting | Default | Description |
|---|---|---|
| Provider | OpenRouter | Any configured provider: Anthropic, OpenAI, Groq, OpenRouter, or Custom |
| Model | openai/gpt-4o-mini |
Model for generating titles |
| Mode | Rename file | Rename file or set frontmatter property |
| Max title length | 8 words | Maximum words in generated title |
Property Filler
| Setting | Default | Description |
|---|---|---|
| Provider | OpenRouter | Any configured provider: Anthropic, OpenAI, Groq, OpenRouter, or Custom |
| Model | openai/gpt-4o-mini |
Model for generating property values |
| Custom prompt | (built-in) | Template with {property} placeholder |
| Overwrite existing | Off | Replace existing property values |
Architecture
NibbleAI/
├── main.ts # Plugin entry point, commands, events
├── src/
│ ├── SettingsManager.ts # Settings + SecretStorage management
│ ├── AIService.ts # Unified AI client (OpenAI-compatible + Anthropic-native)
│ ├── ProviderResolver.ts # Shared provider → API key/endpoint resolution
│ ├── ModelBrowser.ts # OpenRouter model suggest modal
│ ├── PostProcessor.ts # LLM post-processing for transcripts
│ ├── NibbleAISettingsTab.ts # Settings UI
│ ├── recorder/ # Audio recording (Timer, StatusBar, Controls, AudioRecorder)
│ ├── transcription/ # Whisper API calls + template rendering
│ ├── proofreading/ # Proofreading module
│ ├── title/ # Title generation module
│ └── property/ # Property filling module
└── styles.css # Plugin styles
All AI calls go through AIService.callChatCompletion(), which handles both OpenAI-compatible and Anthropic-native API formats. Each feature module is self-contained and receives the plugin instance for access to settings and the Obsidian API.
Development
# Install dependencies
npm install
# Build for production
npm run build
# Watch mode (auto-rebuild on changes)
npm run dev
# Type-check only
npx tsc --noEmit -skipLibCheck
Set the OUTPUT_PATH environment variable to output main.js directly into your vault's plugin directory for fast iteration:
OUTPUT_PATH=/path/to/your/vault/.obsidian/plugins/nibbleai npm run dev
License
MIT
NibbleAI — small bites of AI for your Obsidian vault.