Stated plainly

Security & FAQ

Caroline trades some safety for real autonomy, on purpose. Here's exactly what that means, and how to run her more cautiously if you'd rather.

Why does Caroline need so many permissions?

Because an assistant that has to stop and ask before every mouse click or file read isn't actually saving you time. Broad permissions are the whole point — she can act on a multi-step task without a confirmation dialog at every turn.

Is it safe to give an AI control of my mouse, keyboard, and files?

That's a real trade-off, not a solved problem. Autonomy is prioritized ahead of safety in this project, deliberately. If that's not the trade-off you want on your main machine, run Caroline inside a virtual machine and sandbox her there — fully supported, just not the default.

Does she just act on her own judgment, with no checks at all?

Not for the things that matter most. Broad permissions mean she doesn't ask before every small step — but a separate, independent check still has to agree before she sends anything on your behalf to someone else, or does anything that touches your money. An unanswered yes/no question is treated as a refusal, not a green light to proceed. If that check doesn't find a clear approval already on record, the whole turn stops, the same as pressing Stop yourself. See How She Thinks for exactly how that works.

How are my passwords and API keys handled?

Secrets are never written to local files, chat history, or Skills files. When Caroline needs to remember one, it's saved as a note in a dedicated vault folder instead. Login credentials for connected accounts are never typed into chat at all — the native login window posts straight to the backend over the app's own channel, never through the model's context.

What data leaves my machine, and where does it go?

Chat goes wherever you've pointed it: Anthropic directly through your own Claude Pro/Max subscription or Anthropic Console account, OpenAI through your own ChatGPT or Codex account, or SquirrelWisdom's proxy if you're paying as you go instead — that's how the assistant thinks, whichever one you've chosen. A few optional features (long-term notes, SMS, in-browser document editing, a second-opinion model for high-stakes wording) are backed by a separate SquirrelWisdom account if you choose to connect one; nothing about those is required to use Caroline for chat and desktop automation.

Is this actually open source?

Yes — Apache License 2.0, on GitHub. Fork it, audit it, run it commercially, modify it. The Open Code page lets you browse the current source directly on this site.

What does it cost?

Caroline herself is free. Chatting with her costs whatever your chosen account costs: your own Claude Pro/Max subscription, your own ChatGPT/Codex account, or pay-as-you-go through a SquirrelWisdom wallet if you'd rather not hold a Claude subscription at all — there's no bundled API key either way. The optional SquirrelWisdom-backed features (notes, SMS, document editing) are billed separately if and when you use them.

What platforms does it run on?

Windows 10/11, x64, today. The shell and native automation helpers are Windows-specific; the backend and MCP servers underneath are not. See Contribute if you want to help port her elsewhere.

How do I run Caroline more cautiously?

Install her inside a virtual machine with restricted network/file access, rather than on your main desktop. You keep the same assistant; you just shrink what "acting on your behalf" is able to touch.