Raycast AI Review: What You Get, and Where It Falls Short

Raycast AI is a good buy if you already live in a keyboard launcher and want models, your apps and your MCP servers one hotkey away. It's a weaker buy if you only want a chatbot, or if you hoped your own API key would get you in for free. The Free plan has no AI at all, and that includes bring-your-own-key and local models. Every AI feature sits behind a paid plan, and usage on Raycast's own models is metered in credits.

Here's the short version:

• Best part: AI inside your launcher, with one model picker covering hosted models, your own keys, your Claude or ChatGPT subscription, and Ollama.
• Worst part: credits are hard to predict, and the top-tier models need a pricier plan.
• Platforms: macOS, Windows 10 and later, and iOS, but the feature set isn't identical on all three.

This review is based on Raycast's own pricing page and manual, not on hands-on testing, so it covers documented behaviour rather than how fast or good the answers feel. If you're choosing a launcher rather than judging the AI layer, start with our Raycast vs Alfred comparison.

Quick comparison of Raycast AI features

All of them are marked Pro Exclusive in the manual, which means any paid plan.

What's on which Raycast plan

The plan names trip people up. The manual says Pro, Pro+ and Max. The pricing table shows the same plans as Pro, Plus and Max. The old Advanced AI add-on is gone, and the billing docs say Pro+ replaced it.

Prices and credits are from Raycast's site on November 2, 2026. Yearly billing is cheaper, and team plans cost more per member. One subscription covers all three platforms.

Free is still a serious launcher. It includes the core features, thousands of extensions, custom extensions and Clipboard History with a three-month limit. What it doesn't include is AI, Cloud Sync, Dictation or the Translator.

Two details are easy to miss:

• You can't buy Pro without AI. Raycast's FAQ says so directly. You can switch AI off in Settings, but you'd be paying for it anyway.
• The trial is limited. Raycast says eligible new accounts can start a free Pro trial with Pro models. Advanced models aren't part of it, and credit top-ups aren't available during it.

Quick AI, AI Chat and AI Commands

These three are the daily surface of the product, and they do different jobs.

Quick AI is the reason most people want this at all. You type a question in the launcher, press Tab, and the answer streams into the same window. Press Return and it pastes into the app you were just using. You can keep asking follow-ups, attach the focused window or selected text, and move the whole conversation into AI Chat with a shortcut.

AI Chat is the fuller workspace. It has persistent history, Memory, attachments and tool use, in a window you can keep beside your work. If a question grows into a project, this is where it lives.

AI Commands turn repeat prompts into commands. Built-in ones include Improve Writing, Fix Spelling and Grammar, Find Bugs in Code and Summarize Webpage. You can write your own with a prompt, a model and a creativity setting, and choose where the output goes.

That last point is the most underrated part of the product. A saved command is a small step away from one-off chat: the next time you need that rewrite, you press a key instead of retyping instructions. Each command also has its own model setting, so a cheap model can handle grammar while a stronger one handles harder work.

MCP and AI Extensions

Raycast AI gets more useful when it can reach your tools, and it has two routes.

AI Extensions connect the AI to extensions in the Raycast Store, so you can mention one with @ and ask it to do something. That's an advantage a standalone chatbot doesn't have: the Store already covers many of the apps you'd want to reach.

MCP support covers everything else. If you're new to the term, our guide to what MCP is explains it. On desktop you add a server with an install command, using either a local stdio process or a remote HTTP endpoint. On iOS, servers are HTTP only. Tools from a connected server then work in AI Chat, Quick AI and AI Commands.

The permission default is sensible: Raycast approves read-only tools on its own and asks about the rest. There's also an Always Allow setting, which the manual says to use with caution because it gives the AI full permission over your tools and files. Believe it.

The practical limit is the same as anywhere else. Setting up a stdio server still means commands, arguments and environment variables. Raycast makes installing one easier, but it doesn't make it non-technical.

BYOK, your own subscription and local models

This is where Raycast AI has changed most, and where a lot of older write-ups are out of date. There are now several ways to supply the model yourself:

There's also a Custom Providers option for any OpenAI-compatible endpoint.

Bring Your Own Subscription is the most interesting of these. If you already pay for Claude or ChatGPT, you can use that plan's models in Quick AI, AI Chat, AI Commands and Automations without spending Raycast credits. Your provider's usage limits apply, and they're shared with everything else on that subscription. When you hit one, the chat pauses. Raycast doesn't switch you to credits automatically.

Three catches apply across this section:

1. All of it needs a paid Raycast plan. The manual is explicit that bringing your own AI requires one. So BYOK doesn't remove the Raycast bill, and a Pro subscription on top of a Claude plan is a cost to weigh honestly.
2. BYOK model lists are limited. With Anthropic, Google or OpenAI keys, you only get models Raycast already offers. An OpenRouter key opens OpenRouter's full catalogue.
3. Credits don't fully disappear. A few tools Raycast runs on its own servers, such as image generation, still use a Raycast model and count against your credits. Quick Fix, the one-hotkey grammar correction, always runs on Raycast AI and spends credits, even if you've set AI Commands to use your own key.

Models and credits

Raycast's model list is wide. It spans OpenAI, Anthropic, Google, Perplexity, Mistral, xAI, DeepSeek and others in a single interface. For someone who likes comparing answers, that's a real convenience, and the regenerate-with-another-model shortcut makes it quick.

The trade is the credit system. Raycast describes AI credits as a measure of usage, not a fixed number of messages. Long conversations, attachments, tool calls, and regenerated or edited messages all add up, and each model spends at its own rate.

So you can't translate 500 credits into a number of questions in advance. The model picker does show a Cost meter, and Settings shows usage by model, so you can learn your own rate. But the first month is guesswork.

Other rules worth knowing:

• Unused monthly credits roll over once, and each grant expires 60 days after you receive it.
• Top-ups come in packages and expire after 12 months.
• When credits run out, Raycast AI stops until your next allowance or a top-up. You can keep going with your own key, subscription or local model.
• If you cancel, remaining credits can't be used until you resubscribe.
• Top-tier models, including Claude Opus 5.5 and GPT-6.1 Sol, are limited to Pro+ and Max.

Windows status and platform gaps

Raycast is no longer a Mac product. The Windows page lists v2.10.1.0, installed through WinGet, and the FAQ names Windows 10 21H2 and later plus Windows 11. The manual marks Quick AI, AI Chat, AI Commands, MCP, BYOK, Local Models, Agents and Screen Awareness as available on Windows.

It isn't full parity yet. Raycast's own page says it's still working on bringing all the macOS features to Windows, and the manual's platform badges for built-in AI Extensions still differ. Extensions that need native code won't necessarily work on both platforms.

Mac users have a different thing to check. Raycast's AI page says "macOS Tahoe and Apple Silicon required" beside the download button. If you're on an Intel Mac or an older macOS, confirm what you can run before paying.

Privacy

The documentation here is detailed, which is a point in Raycast's favour. Per its privacy page:

• Raycast says it doesn't log or retain user prompts, and that your data isn't used to train models.
• Models provided through Raycast are zero data retention, with three exceptions: Mistral, xAI and Claude Fable 5.1, which keep data for 30 days.
• With Cloud Sync on, AI chats are stored on Raycast's servers, encrypted.
• Local model conversations run on your machine, apart from the server-side tools noted above.
• Even when you supply the model, Raycast says it still stores your chat history and syncs it if Cloud Sync is on. Attachments upload by default, and images only stay on your device with Cloud Sync or AI Chat History sync off.
• If AI Memory is on, recent chats from any model are processed by a Raycast-hosted model.

If you connect your own provider, that provider's terms apply instead. For sensitive work, the local route is the one that keeps prompts away from model providers, but check Cloud Sync and Memory too.

Where Raycast AI falls short

Pulling the weak spots together:

• No free way in. Plenty of tools let you use your own key at no charge. Raycast doesn't.
• Credit predictability. Usage-based pricing is fair, but you can't budget it before you've used it.
• Plan gating. The cheapest paid plan leaves out the top-tier models, which may be the ones you came for.
• A lot of surface. Quick AI, Chat, Commands, Agents, Skills, Projects, Automations and Extensions is many concepts for someone who wanted a faster way to ask a question.
• Setup still exists. MCP servers and Bring Your Own Subscription both assume comfort with a terminal or config details.
• Stays close to the launcher. It's strongest for quick questions and small actions. Automations are device-local, and permission prompts pause tool use by default.

None of these is a dealbreaker on its own, but the first two are the ones to weigh before you subscribe. If they rule it out, our Raycast alternatives guide separates the launcher from the AI layer.

Who Raycast AI suits

Good fit:

• You already use Raycast and keep leaving it to open a chatbot.
• You pay for Claude or ChatGPT and want it in your launcher.
• You switch between macOS and Windows and want one setup.
• You run repeat prompts all day and want them on keys.

Poor fit:

• You only want a chatbot. The provider's own app is simpler.
• You want AI with no subscription. Free won't do it, even with your own key.
• You need heavy usage of top-tier models at a low, fixed cost.

FAQ

Is Raycast AI free?

No. Raycast says AI is included in paid plans and isn't available on Free. Eligible new accounts can start a free Pro trial.

Is Raycast AI worth it?

If you use the launcher daily, it's reasonable, because the value is reach, not model quality. The same models exist elsewhere. You're paying to have them in the launcher, connected to your extensions and commands.

Can I use my own API key with Raycast AI?

Yes, on any paid plan, with Anthropic, Google, OpenAI or OpenRouter keys. You pay the provider directly, on top of your Raycast plan.

Does Raycast AI work on Windows?

Yes. The manual lists Quick AI, AI Chat, AI Commands, MCP and BYOK on Windows, though Raycast says it's still bringing over other macOS features.

Does Raycast AI support MCP?

Yes, on Mac, Windows and iOS. Desktop supports stdio and HTTP servers, and iOS supports HTTP only.

Can Raycast AI run models locally?

Yes, through Ollama on Mac and Windows, on a paid plan. The conversation runs on your machine, though a few server-side tools still go online, and Cloud Sync and AI Memory have their own rules.

Key points

• Raycast AI needs a paid plan, even for your own key or local models.
• Quick AI and AI Commands are the strongest reasons to buy, because they bring AI into the launcher.
• Bring Your Own Subscription is the best option if you already pay for Claude or ChatGPT.
• Credits vary by model and task, so expect a learning month.
• Windows covers the main AI features, but full parity isn't there yet.

Raycast AI is built for quick actions from a launcher. If you'd rather run a complete AI app or workflow than a one-line prompt, Taku is a different shape: an AI-native desktop workspace where you mirror a setup someone else already got working, then save it and run it again. It isn't a launcher, so it sits alongside one. Taku is in Beta, and the Mac app is available now.