Cai

Select any text. Get smart actions.

A privacy-first clipboard assistant powered by local AI.
Your data never leaves your machine.

getcai.app

---

Cai is a native macOS menu bar app that detects what's on your clipboard and offers smart, context-aware actions. Copy a meeting invite and it creates a calendar event. Copy an address and it opens Maps. Copy any text and ask your local AI to summarize, translate, or do anything you want — all without leaving your keyboard.

No cloud. No telemetry. No accounts.

How It Works

1. Select text anywhere on your Mac
2. Press ⌥C (Option+C)
3. Cai detects the content type and shows relevant actions
4. Pick an action with arrow keys or ⌘1–9
5. Result is auto-copied to your clipboard — just ⌘V to paste

Examples:

• Select "serendipity" → Define, Explain, Translate, Search
• Select "Let's meet Tuesday at 3pm at Starbucks" → Create calendar event, Open in Maps
• Select "123 Main St, NYC 10001" → Open in Maps
• Select an email in Mail → Reply, Summarize, Translate
• Select a Slack message → Reply (matches the conversation tone)
• Select a recipe → Custom Action: "Extract ingredients for 2 people"
• Select meeting notes → Shortcut: "Format as action items with owners and deadlines"
• Select a project name → Shortcut: search in Google Drive (https://drive.google.com/drive/search?q=%s)
• Select {"name": "John"} → Pretty Print JSON

Features

• Smart detection of 7 content types with context-aware actions
• Built-in AI — ships with Ministral 3B, downloaded on first launch. Also works with LM Studio, Ollama, or any OpenAI-compatible server
• Custom AI action (⌘1) — set your own prompt to do anything: improve writing, create email replies, translate, count words
• Custom shortcuts — save reusable prompts and URL templates, access them by typing to filter
• Output destinations — send results to Mail, Notes, Reminders, or custom webhooks, URL schemes, AppleScript, and shell commands
• Type-to-filter — start typing to filter actions, shortcuts, and destinations by name
• Clipboard history — access last 9 items with ⌘0
• App-aware — Cai knows which app you're in (Mail, Slack, Safari…) and adapts AI responses to match the context
• Keyboard-first — navigate and execute everything without touching the mouse
• Privacy-first — no internet required, no data leaves your machine

Content Types & Actions

Content Type	Detection	Actions
Word	1–2 words	Define, Explain, Translate, Search
Short Text	< 100 characters	Explain, Reply, Proofread, Translate, Search
Long Text	100+ characters	Summarize, Explain, Reply, Proofread, Translate
Meeting	Date/time references (short text only)	Create Event, Open in Maps + all text actions
Address	Street patterns, "at [Place Name]" (short text only)	Open in Maps + all text actions
URL	https:// or www.	Open in Browser (+ text actions if URL has surrounding text)
JSON	Valid JSON object/array	Pretty Print

All types also get Custom Action (⌘1) for free-form AI prompts. Meeting and address detection is limited to short text (≤200 chars) — longer text is treated as text so you always have access to Summarize, Reply, etc.

Keyboard Shortcuts

Shortcut	Action
⌥C	Trigger Cai (global hotkey)
↑ ↓	Navigate actions
↵	Execute selected action
⌘1–9	Jump to action by number
⌘0	Open clipboard history
⌘↵	Submit custom prompt / Copy result
A–Z	Type to filter actions and shortcuts
Esc	Clear filter / Back / Dismiss

Installation

Download

1. Download the .dmg from the latest release
2. Open the DMG and drag Cai.app to your Applications folder
3. Open the app and grant Accessibility permission (see below)
4. Cai will download a small AI model (~2 GB) on first launch — or skip if you already use LM Studio / Ollama

First Launch Setup

On first launch, Cai needs Accessibility permission to use the global hotkey (⌥C) and simulate ⌘C to copy your selection.

Step 1 — Open Cai. It will ask for Accessibility permission. Click Open System Settings.

Step 2 — Toggle Cai on in the Accessibility list.

You're all set! Press ⌥C with any text selected to start using Cai.

Build from Source

git clone https://github.com/soyasis/cai.git
cd cai/Cai
open Cai.xcodeproj

In Xcode:

1. Select the Cai scheme and My Mac as destination
2. Product → Run (⌘R)

Note: The app requires Accessibility permission and runs without App Sandbox (required for global hotkey and CGEvent posting).

LLM Setup

Built-in (zero config)

Cai ships with a bundled AI engine (llama.cpp). On first launch it downloads Ministral 3B (~2 GB) and runs everything locally — no external server needed.

The model is stored in ~/Library/Application Support/Cai/models/. The engine starts automatically on launch and stops when you quit Cai.

External providers

If you already use LM Studio, Ollama, or another local server, skip the built-in download. Cai works with any OpenAI-compatible endpoint.

Provider	Default URL	Setup
LM Studio	http://127.0.0.1:1234/v1	Download → Load model → Start server
Ollama	http://127.0.0.1:11434/v1	Install → ollama pull qwen3:4b
Jan AI	http://127.0.0.1:1337/v1	Download → Load model → Start server
LocalAI	http://127.0.0.1:8080/v1	Setup guide
Open WebUI	http://127.0.0.1:8080/v1	Install → Enable OpenAI API
GPT4All	http://127.0.0.1:4891/v1	Download → Enable API server
Custom	User-defined	Any OpenAI-compatible server

Auto-detection: On launch, Cai checks if the current provider is reachable. If not, it probes known local ports and switches to the first one that responds — no manual setup needed.

To configure manually: Open Cai Preferences (left-click menu bar icon) → select your Model Provider.

AI is optional — system actions (Open URL, Maps, Calendar, Search, Pretty Print JSON) work without it.

Recommended Models

The built-in model (Ministral 3B) is a great starting point — it's fast, concise, and works out of the box with no setup. For most users, there's no need to change it.

If you want to experiment with different models or want more control, you can use an external provider like LM Studio or Ollama. These are aimed at advanced users comfortable with managing local AI servers.

Model	Params	Why	Install
Ministral 3B (built-in)	3B	Fastest real-world feel — concise output, clean markdown, fewest tokens per answer. Ships with Cai.	Automatic on first launch
Qwen3 4B	4B	Smartest at this size — best on STEM, coding, and instruction following. Can be verbose.	lms get qwen3-4b / ollama pull qwen3:4b
Llama 3.2 3B	3B	Reliable all-rounder — solid summarization, Q&A, and general conversation.	lms get llama-3.2-3b-instruct / ollama pull llama3.2:3b
Gemma 3 4B	4B	Strong multilingual support (140+ languages).	lms get gemma-3-4b / ollama pull gemma3:4b
Qwen3 8B	8B	Best quality in the small range. Slower than 3–4B.	lms get qwen3-8b / ollama pull qwen3:8b

Tip: Switch to Qwen3 4B if you need more intelligence (e.g., complex instructions or coding tasks).

LM Studio vs Ollama: Both work great with Cai. LM Studio tends to be noticeably faster for inference (aggressive Metal GPU acceleration, speculative decoding), which matters for a clipboard tool where you want sub-second responses. Ollama is simpler to set up and manage models via CLI. Try both and see what feels snappier on your hardware.

Configuration

Left-click the Cai menu bar icon (or click the logo in the action window footer) to access Preferences:

Setting	Description	Default
Translation Language	Target language for translations	English
Search URL	Base URL for web searches	Brave Search
Maps Provider	Apple Maps or Google Maps	Apple Maps
Model Provider	Built-in, LM Studio, Ollama, or Custom	Built-in (auto-detected)
Custom Action	Free-form AI prompt via ⌘1	—
Custom Shortcuts	Save prompt and URL shortcuts for instant access	—
Output Destinations	Where to send results (Mail, Notes, webhooks, etc.)	Email + Notes enabled
Launch at Login	Start Cai automatically	On

Custom Shortcuts

Save frequently used prompts and URL templates as shortcuts. They appear when you type to filter the action list.

Two types:

• Prompt — sends your clipboard text + saved prompt to the local LLM (e.g., "Rewrite as email reply", "Convert to CSV")
• URL — opens a URL with your clipboard text substituted via %s (e.g., https://reddit.com/search/?q=%s)

To create: Preferences → Custom Shortcuts → click + → add a name, pick the type, and enter the prompt or URL template.

To use: Press ⌥C, then start typing the shortcut name. Matching shortcuts appear alongside filtered built-in actions.

Output Destinations

After an AI action processes your text (or directly from the action list), you can send the result to an output destination instead of just copying it to the clipboard.

Built-in Destinations

Destination	What it does	Default
Email	Opens Mail.app with a new draft containing the text	Enabled
Save to Notes	Creates a new note in Notes.app (preserves line breaks)	Enabled
Create Reminder	Adds a reminder to your default Reminders list	Disabled

Toggle built-in destinations on/off in Preferences → Output Destinations.

Custom Destinations

Create your own destinations to send text to any app or service:

Type	Use case	Example
Webhook	Send to any API via HTTP POST/PUT/PATCH	Post to Slack channel, create Notion page
AppleScript	Control any macOS app	Add to Things, create OmniFocus task
URL Scheme	Open deep links with your text	Save to Bear, open in Obsidian
Shell Command	Run terminal commands	gh issue create, pipe to a script

Template placeholders:

• {{result}} — your clipboard or AI-processed text (auto-escaped for the destination type)
• {{field_key}} — value from a setup field (e.g. {{api_key}} for API tokens)

Setup fields let you store secrets like API keys. They're resolved at execution time and masked in the UI.

"Show in action list" — enable this to make a destination appear as a direct action (skips the AI step). Useful for quick-send workflows like "Send to Slack" directly from the action list.

To create: Preferences → Output Destinations → click + → pick a type, configure the template, and optionally add setup fields.

Requirements

• macOS 13.0 (Ventura) or later
• Apple Silicon (M1 or later) for the built-in AI engine
• ~2.5 GB disk space for the bundled model (downloaded on first launch)
• Accessibility permission (for global hotkey ⌥C)

Troubleshooting

macOS blocks Cai from opening (building from source) If you build from source without a Developer ID, macOS Gatekeeper may block the app. Remove the quarantine flag via Terminal:

xattr -cr /Applications/Cai.app

This is not needed when installing from the official DMG release.

Global shortcut ⌥C doesn't work

• Check System Settings → Privacy & Security → Accessibility — make sure Cai is listed and enabled
• If it's listed but still not working, remove Cai from the list and re-add it
• Make sure no other app is using ⌥C (e.g., Raycast, Alfred, BetterTouchTool)

LLM not connecting

• Verify your server is running: curl http://127.0.0.1:1234/v1/models
• Check that the URL in Preferences matches your server's address and port
• Ollama uses port 11434, LM Studio uses 1234 — make sure you selected the right provider

Date/meeting not detected

• Detection works best with English dates ("Tuesday at 3pm", "lunch tomorrow at noon")
• Try rephrasing: "3pm tomorrow" instead of "tomorrow 15h"
• Some informal formats may not be recognized — explicit dates work most reliably

Tech Stack

• SwiftUI + AppKit (native macOS, no Electron)
• CGEvent for Cmd+C simulation (private event source to isolate modifier state)
• AXUIElement for text selection detection
• NSPanel subclass for floating window that captures keyboard events
• Bundled llama.cpp for local LLM inference (ARM64 macOS, Metal GPU)
• Actor-based LLMService, BuiltInLLM, and OutputDestinationService for thread-safe async/await
• ICS file generation for calendar events (no EventKit permissions needed)
• AppleScript integration for native app destinations (Mail, Notes, Reminders)
• SMAppService for Launch at Login
• HotKey (SPM) for global keyboard shortcut

License

MIT