Select any text or image. Run custom AI actions, chain them into pipelines, and route results anywhere. Zero app switching.
Website · Docs · Download · Extensions
Select any text or image anywhere, press ⌥C, and run AI prompts, shell scripts, and connectors like GitHub and Linear. Chain them into pipelines that fire on a single keystroke. Without the app switching. One hotkey, no config, everything runs locally.
No cloud. No telemetry. No accounts.
- Select text or copy an image anywhere on your Mac
- Press ⌥C (Option+C)
- Cai detects the content type and shows relevant actions
- Pick an action with arrow keys or ⌘1–9
- Hit return to finish. Result is auto-copied to your clipboard — just ⌘V to paste
Examples:
- Take a screenshot → Create GitHub issue
- Select a recipe → Ask AI: "Extract ingredients for 2 people"
- Select
"serendipity"→ Define, Explain, Translate, Search - Select
"Let's meet Tuesday at 3pm at Starbucks"→ Create calendar event, Open in Maps - Select an email in Mail → Reply, Summarize, Translate
Chain example: Select an article URL → fetch the page → summarize → post to Slack. Four steps. One ⌥C.
→ Read the full How It Works guide
Cai speaks MCP, so Claude Code, Cursor, or Codex can author your actions for you. Describe what you want in the agent you already use and it proposes an action. Cai holds the proposal until you approve it, and from then on the action runs from ⌥C on its own, with no agent in the loop. Shell and prompt actions run locally by default.
In Cai, open Settings → MCP → Server and copy the line for your agent. For Claude Code:
claude mcp add --scope user cai -- ~/Library/Application\ Support/Cai/bin/cai-mcpYour agent gets four tools: list_actions, get_action, create_action, update_action. There is no tool that runs an action, deletes one, or approves a proposal, so the most a confused or hostile agent can achieve is a proposal you read and refuse. Cai shows you the whole action before anything is saved, and flags the parts that deserve a second look: running a shell command, putting your selection into a URL, or pasting over your selection without showing you first.
The bridge is a small helper that your agent launches itself and talks to over stdin and stdout. There is no port and no listener, so nothing else on your Mac or your network can reach it.
- Smart content detection — recognizes what you copied (text, image, URL, JSON, meeting, address) and shows the right actions
- Built-in AI — Apple Intelligence on macOS 26+, or in-process MLX inference on Apple Silicon. No server, no cloud, no setup
- GitHub & Linear — create issues from any selected text with AI-generated title, body, and duplicate detection
- Custom actions — save reusable AI prompts, URL templates, and shell commands as one-click actions
- Agent-authored actions — let Claude Code, Cursor, or Codex write your actions over MCP, approve them once in Cai, run them from ⌥C forever (see above)
- Action chains — pipe selection through AI prompts, scripts, and destinations (Slack, GitHub, Apple Shortcuts) in one keystroke. Save the chain once, run it forever.
- Image to Text — on-device OCR via Apple Vision framework
- Bring your own LLM — works with LM Studio, Ollama, any OpenAI-compatible server, or any model from HuggingFace mlx-community
Also includes:
- Custom output destinations (Mail, Notes, webhooks, AppleScript)
- Follow-up questions
- Context Snippets (pass per-app context to the LLM)
- Clipboard history (last 100, search and pin)
- Keyboard-first (arrow keys, ⌘1–9)
- Community extensions
→ See all features in the docs
⚠️ One-time manual update needed. We rotated Cai's update-signing key, so 1.5.1 and older can't auto-update to 1.5.2. Download it from the latest release or getcai.app once; auto-updates resume from then on. (A brief "improperly signed" warning is expected and safe.)
brew tap cai-layer/cai && brew install --cask cai- Download the
.dmgfrom the latest release - Open the DMG and drag Cai.app to your Applications folder
- Open the app and grant Accessibility permission when prompted
- On macOS 26+, Cai uses Apple Intelligence automatically. Otherwise, the built-in MLX model downloads on first launch — or skip if you already use LM Studio / Ollama
→ Full installation guide · LLM setup
git clone https://github.com/cai-layer/cai.git
cd cai/Cai
open Cai.xcodeprojIn Xcode: select the Cai scheme and My Mac as destination, then Product → Run (⌘R).
Note: The app requires Accessibility permission and runs without App Sandbox (required for global hotkey and CGEvent posting).
→ Check the full changelog
Full documentation is at getcai.app/docs:
- How It Works — content detection, smart actions, follow-ups
- Keyboard Shortcuts — every key and what it does
- LLM Setup — Apple Intelligence, MLX, LM Studio, Ollama, cloud providers
- Choosing a Model — model picker guide and quantization explainer
- Ask AI — free-form prompts on selected text
- Custom Actions — save prompts, URLs, and shell commands
- Action Chains — chain custom actions, destinations, and Apple Shortcuts into one-keystroke pipelines
- Custom Destinations — webhooks, AppleScript, deeplinks, shell
- Connectors — GitHub and Linear integration
- Context Snippets — per-app context for smarter actions
- Community Extensions — install and create shared actions
- Troubleshooting — common issues and fixes
- macOS 14.0 (Sonoma) or later
- Apple Silicon (M1 or later) for the built-in AI engine
- 8 GB RAM minimum, 16 GB recommended for larger models
- Accessibility permission (for global hotkey ⌥C)
- SwiftUI + AppKit — native macOS, no Electron
- MLX-Swift — in-process LLM inference on Apple Silicon, no subprocess or server
- No App Sandbox — global hotkey requires CGEvent posting outside the sandbox
- MCP via ~200-line JSON-RPC client (Beta) — GitHub and Linear connectors with zero external MCP dependencies