Built for how you actually work
Pick your role
The same assistant, pointed at very different workdays.
Caroline runs as a tray app with a global hotkey and a real chat window — no terminal, no IDE, nothing to configure before she's useful.
Reads and sends real email over IMAP/SMTP, keeps long-term notes, opens and edits docx/xlsx/pptx/pdf files in a real embedded editor.
Sees your screen, moves the mouse, types, drives a real browser session — and can text or check messages from your own number through a paired Android phone.
Talks back on demand with low-latency local TTS, listens through a mic button, and can show an animated on-screen presence instead of a wall of text.
Long-running sessions don't get cut off or blow their context — old context ages out automatically while she keeps working, indefinitely.
A hung or crashed session gets caught by a watchdog and restarted automatically, replaying whatever didn't get answered. You never have to notice.
Looks things up on Wikipedia, pulls out sourced facts, and searches across documents you hand her — with a citation, not a guess.
Everything the desktop app can do is also a real HTTP API — reachable from another device, locked to your own account. No shared backdoor, no separate password. Read the API guide →
An unanswered yes/no question is treated as a no, not a green light. Anything that messages someone else on your behalf, or touches your money, needs a real answer on record first — otherwise the whole turn stops. See how she checks herself →
Pick whichever account you already have — switch any time, per conversation, with nothing to migrate.
Log in with a Claude Pro/Max subscription or an Anthropic Console account. Full control, billed the way you already bill Anthropic.
No Claude account at all — top up a wallet and go. The simplest way to try Caroline today if you don't already have a Claude subscription.
Sign in with ChatGPT or a Codex API key and Caroline runs on OpenAI's models instead — same tools, same chat window, same assistant.
"Caroline isn't a tool. She's an assistant — with her own opinion, and her own responsibility for the outcome."
"She holds broad permissions by design, because she has to be able to act on her own. Autonomy comes first — deliberately, ahead of safety. If that trade-off isn't for you, run her in a VM and sandbox her there."
The same assistant, pointed at very different workdays.
Caroline is Apache-2.0 licensed and built to be extended. Add your own MCP servers, port her to a new platform, wire in a tool she doesn't have yet. This is AI for everyone, built in the open — not a product roadmap you wait on.
See how to contribute| Shell | WPF (C#), tray app, global hotkey |
|---|---|
| Backend | Python (FastAPI), Claude Agent SDK |
| Chat account | Your own Claude, pay-as-you-go via SquirrelWisdom, or your own ChatGPT |
| Tools | Standard MCP servers, add/remove freely |
| License | Apache 2.0 |
Requires Windows 10/11 (x64) and the WebView2 runtime (usually already installed). No Claude subscription required to get started — bring your own Claude or ChatGPT account, or pay as you go through SquirrelWisdom.
