Why you can trust what's running on your machine
AI automation asks a lot of trust. We try to earn it not by making promises, but by designing the software so that trusting it isn't necessary — you can verify everything it does, and prevent anything you don't want.
Local-first by architecture
Your memories, preferences, chat history, and task receipts are stored on your own disk — inside a WAL-mode SQLite vault that never talks to a server we control. There is no "our cloud" holding copies of your data, because there is no cloud on our side.
When Ghosthread needs to reason about something complex, it sends the specific minimal context to an AI provider you choose. Everything else — your files, your screen, your clipboard, your history — stays put.
You govern the permissions
Ghosthread ships with an Access Control layer that lets you toggle entire categories of capabilities on or off:
- • Shell command execution
- • Filesystem writes and reorganization
- • Clipboard access
- • Any specific app or folder you want off-limits
When you turn something off, it's off at the executor level — the AI cannot argue its way around it. Restrictions are persisted and survive restarts.
Every action has a receipt
Whenever Ghosthread does something, it writes a structured record into your local vault: which steps ran, what each step touched, whether it succeeded, and how long it took. Nothing is hidden.
You can inspect the receipt at any time from the dashboard. This isn't a log for debugging — it's a transparency mechanism. If you ever wonder "what did it just do?", the answer is one click away.
Safety that survives the AI being wrong
AI models can be wrong, confused, or manipulated. Ghosthread assumes this and enforces its hardest safety rules at the executor layer — the last line of code before the operating system is touched. Even if the AI tries to do something dangerous, the executor refuses.
This is the difference between "trust the AI to be careful" and "the AI literally cannot touch your C:\Windows folder." We built the second kind. See the Security page for the full five-layer model.
What we don't do — as a company, and as a piece of software
- ✗ We don't sell your data. Ever.
- ✗ We don't collect your files or screenshots.
- ✗ We don't profile you or run ad targeting.
- ✗ We don't inject telemetry silently.
- ✗ The software doesn't modify system folders.
- ✗ The software doesn't permanently delete files.
- ✗ The software doesn't act autonomously without permission.
- ✗ The software doesn't operate during games or anti-cheat sessions.