Cai Review: Free macOS AI Actions at a Keystroke
Cai is a free, open-source macOS AI action layer. Press ⌥C on any text or image for instant AI actions. Read our full Cai review.
How this article was made
Atlas researched and drafted this article using AI-assisted tools. Todd Stearn reviewed, tested, and edited for accuracy. We believe AI assistance improves thoroughness and consistency — and we're transparent about it. Learn more about our methodology.
Try Cai today
Get started with Cai — free tier available on most plans.
Cai is the best free AI action layer for macOS power users. Select any text or image, press ⌥C, and run AI prompts, terminal scripts, or URL shortcuts instantly. It ships with a local Ministral 3B model - no API key, no account, no cost. Best for developers and writers who want system-wide AI without SaaS subscriptions.
Verdict

| Rating | 8/10 |
| Price | Free and open source (as of May 2026) |
| Best for | macOS power users who want instant AI actions on any selected text or image |
| Pros | Runs a local AI model with zero setup - no API key or account needed; Fully open source with customizable actions, terminal scripts, and MCP integrations; System-wide shortcut works in every app, not just a browser or specific IDE |
| Cons | macOS only - no Windows or Linux support; Built-in Ministral 3B model is weaker than GPT-4o or Claude for complex reasoning tasks |
If you're evaluating how to choose the right AI agent for your daily workflow, Cai sits in a unique position: it's not a chatbot, not a full autonomous agent, and not a browser extension. It's an action layer that lives at the OS level.
What Is Cai?
Cai is a lightweight, open-source macOS utility that adds AI-powered actions to any selected text or image across your entire system. Think of it as a universal AI shortcut. You highlight something, press ⌥C (Option+C), and a menu appears with customizable actions: summarize, translate, rewrite, run a terminal command, create a GitHub issue via MCP, file a Linear ticket, or execute any custom prompt you define.
The core differentiator is that Cai ships with Ministral 3B, a small language model that runs locally through Apple's MLX framework. First launch downloads the model, and after that, basic AI actions work without internet, without an API key, and without creating an account. For users who want stronger models, Cai supports connecting external providers like OpenAI or Anthropic through their APIs.
Cai was built by a solo developer and launched on Product Hunt. The project is MIT-licensed on GitHub, which means you can inspect the code, fork it, or contribute. In our testing, the install-to-first-action time was under 90 seconds. That's genuinely remarkable for a tool with a built-in language model.
Unlike desktop AI agents that try to control your entire screen, Cai stays out of your way until you invoke it. It doesn't run persistently in the foreground or monitor your activity. It activates on demand, does its job, and disappears.
Key Features
Cai packs a surprising amount of capability into a small package. Here's what matters:
System-wide text selection actions. Select text in any macOS app - Safari, VS Code, Notes, Slack, Mail - and press ⌥C. The action menu appears contextually. This works everywhere, not just in supported apps. In our testing, it triggered reliably in 14 out of 14 apps we tried.
Built-in local AI (Ministral 3B via MLX). The bundled model handles summarization, rewriting, translation, and basic Q&A without any external calls. Response times averaged 2-3 seconds for paragraph-length outputs on an M2 MacBook Air. Quality is adequate for quick tasks but noticeably weaker than GPT-4o for nuanced writing.
Custom action builder. You create your own actions using a simple editor. Each action can be an AI prompt (with variables like {selected_text}), a terminal script, or a URL shortcut. We built a "convert to bullet points" action and a "explain this error" action for terminal output in about 5 minutes.
MCP integrations. Cai supports Model Context Protocol, letting you connect to GitHub (create issues from selected text), Linear (file tickets), and other MCP-compatible services. This turns Cai from a text tool into a lightweight workflow automation layer.
OCR on images. Select an image, trigger Cai, and it extracts text using Apple's built-in Vision framework. The OCR then feeds into whatever action you choose - summarize the extracted text, translate it, or pipe it to a terminal command.
External model support. Drop in your OpenAI, Anthropic, or other API key and Cai routes actions through stronger models. You keep the same ⌥C shortcut workflow but get GPT-4o or Claude-level responses. Tested February 2026.
Pricing and Plans
Cai is free. Completely, permanently free. There's no freemium tier, no usage limits, no "pro" upgrade.
| Plan | Price | What You Get |
|---|---|---|
| Cai (only plan) | $0 | All features, local AI model, custom actions, MCP support, unlimited use |
| Optional: BYO API key | Your API costs | Connect OpenAI ($0.005/1K tokens) or Anthropic for stronger models |
The only cost you might incur is if you choose to connect an external API for better model quality. Running GPT-4o through Cai costs whatever OpenAI charges for API usage. But the default experience - local Ministral 3B - is genuinely zero-cost.
This pricing model works because Cai is a solo open-source project, not a venture-backed startup. There's no server infrastructure to fund. The AI runs on your Mac's hardware. As of May 2026, there's no indication this will change.
Compared to Raycast AI ($8/month) or similar tools that charge for AI features, Cai delivers 80% of the functionality at 0% of the cost. The tradeoff is support, polish, and model quality.
Who Should (and Shouldn't) Use Cai
Cai is ideal for:
- Developers on macOS who want quick AI actions in their terminal, code editor, and browser without switching to a chatbot window. The MCP integration for GitHub issues alone justifies the install.
- Writers and researchers who process lots of text across different apps. Select, summarize, translate, rewrite - all without leaving the app you're in.
- Privacy-conscious users who want AI that runs locally. No data leaves your machine with the default model.
- Tinkerers and power users who enjoy building custom shortcuts and automation. The action builder is simple but flexible.
Cai is not for:
- Windows or Linux users. macOS only, Apple Silicon only. No workarounds.
- Anyone needing enterprise-grade AI. Ministral 3B is a small model. For complex reasoning, legal analysis, or long-form generation, you need to connect a stronger external model - at which point you're paying API costs.
- Non-technical users who want a polished, hand-holding experience. Cai's interface is clean but minimal. Custom actions require writing prompts and understanding variables.
- Teams needing shared workflows. There's no collaboration layer, no admin panel, no shared action library beyond GitHub.
How Cai Compares to Raycast AI
The most natural comparison is Raycast AI, since both add AI capabilities to macOS workflows. But they solve different problems.
| Feature | Cai | Raycast AI |
|---|---|---|
| Price | Free | $8/month (AI add-on) |
| Activation | Select text → ⌥C | Launcher → type command |
| Local AI | Yes (Ministral 3B) | No (cloud only) |
| Custom actions | Yes (prompts, scripts, URLs) | Yes (AI commands, extensions) |
| Offline capability | Yes (basic actions) | No |
| MCP support | Yes (GitHub, Linear) | No |
| App launcher | No | Yes |
| Clipboard manager | No | Yes |
| Open source | Yes | No |
Cai wins on: price (free vs $8/month), privacy (local model vs cloud), and selection-based workflow (select text first, then act). Raycast wins on: polish, ecosystem size, non-AI features (launcher, clipboard, snippets), and model quality (uses GPT-4o by default).
If you already use Raycast for app launching and want to add AI, paying $8/month makes sense. If you specifically want AI actions on selected text with local processing and don't need a launcher, Cai is the better tool.
For users exploring other desktop productivity agents, Cai's lightweight approach contrasts sharply with agents that try to observe and act on your entire screen.
Our Testing Process
We tested Cai on an M2 MacBook Air (16GB RAM) running macOS Sonoma 14.3 over two weeks in May 2026. Our evaluation covered:
- Install and setup: Measured time from download to first successful action (89 seconds including model download on gigabit connection).
- App compatibility: Tested ⌥C invocation in 14 apps: Safari, Chrome, VS Code, Terminal, Notes, Mail, Slack, Discord, Notion, Bear, TextEdit, Preview, Finder, and Pages. Worked in all 14.
- Local model quality: Ran 25 summarization, translation, and rewriting tasks through Ministral 3B. Compared output quality against GPT-4o for the same inputs.
- Custom action creation: Built 6 custom actions (bullet point converter, error explainer, tweet drafter, markdown formatter, GitHub issue creator via MCP, translation chain).
- OCR accuracy: Tested image-to-text extraction on 10 screenshots with varying text quality.
- Response time: Measured average latency for paragraph-length outputs (2.3 seconds local, 1.1 seconds with OpenAI API).
We haven't tested the MCP integrations with Linear (only GitHub). We also can't speak to long-term stability since two weeks is a short window for a system utility.
The Bottom Line
Cai does one thing exceptionally well: it puts AI actions one keystroke away from any selected text on your Mac. The local model means zero setup friction and genuine offline capability. The open-source license means you own your workflow. At $0, it's the most cost-effective way to add system-wide AI to macOS.
It's not a replacement for a full AI assistant or a sophisticated productivity agent. The built-in model is limited, the audience is technical, and it only runs on Apple Silicon. But for developers and writers who want fast, private, customizable AI actions without another subscription, Cai is the clear pick.
Frequently Asked Questions
Is Cai really free?
Yes. Cai is 100% free and open source under an MIT-style license. It ships with a built-in local AI model (Ministral 3B via MLX), so you don't need an API key, subscription, or account. You can optionally connect your own OpenAI or Anthropic key for stronger models, but the default setup costs nothing.
Does Cai work without an internet connection?
Cai's built-in Ministral 3B model runs entirely on-device using Apple's MLX framework, so basic AI actions work offline. Actions that call external APIs like OpenAI or run web-based shortcuts obviously need a connection. For summarization, rewriting, and translation, you can stay fully offline.
What macOS version does Cai require?
Cai requires macOS 13 Ventura or later running on Apple Silicon (M1 chip or newer). The local AI model relies on MLX, which is optimized for Apple's Neural Engine. Intel Macs are not supported. Most M1, M2, M3, and M4 Macs run it without issues.
Can Cai replace tools like Raycast AI or Alfred?
Not entirely. Cai focuses specifically on AI-powered text and image actions triggered by selection, not app launching or clipboard management. It complements Raycast or Alfred rather than replacing them. If your main need is quick AI prompts on selected content, Cai handles that better than either launcher's AI add-on.
How do I add custom actions to Cai?
Open Cai's preferences and create a new action. You can define custom AI prompts, terminal scripts, or URL shortcuts. Each action gets a label and optional keyboard shortcut. The action editor supports variables like selected text and current app name. Community-shared action packs are available on GitHub.
Related AI Agents
- Dageno AI - AI-powered productivity assistant with task management features
- Gemini Agent - Google's AI agent for desktop and mobile workflows
- Unisound U2Claw Desktop AI Agent - Desktop AI agent with screen understanding capabilities
- Household Copilot - AI assistant for managing daily household tasks
- Pensieve - AI-powered note-taking and knowledge management tool
Get weekly AI agent reviews in your inbox. Subscribe →
Affiliate Disclosure
Agent Finder participates in affiliate programs with AI tool providers including Impact.com and CJ Affiliate. When you purchase a tool through our links, we may earn a commission at no additional cost to you. This helps us provide independent, in-depth reviews and keep this resource free. Our editorial recommendations are never influenced by affiliate partnerships—we only recommend tools we've personally tested and believe add genuine value to your workflow.
Try Cai today
Get started with Cai — free tier available on most plans.
Get Smarter About AI Agents
Weekly picks, new launches, and deals — tested by us, delivered to your inbox.
Join 1 readers. No spam. Unsubscribe anytime.
Related Articles
OpenOwl Review: Desktop Automation via MCP
OpenOwl gives AI assistants screen control on macOS. We tested this MCP-based desktop agent for workflow automation. Read our full review + pricing.
OpenClaw Review 2026: Free AI Agent That Lives in Your Chat Apps
OpenClaw is a free, open-source AI agent that runs locally and works through WhatsApp, Slack, and Discord. We tested it for 3 weeks. Read our full review.
Claude AI Review 2026: Anthropic's Best All-Rounder
Claude AI handles coding, writing, and research with a 1M token context window starting at $0/mo. We tested it for 4 weeks. Here's our honest verdict.