Frequently asked questions

Yes — genuinely private, not “trust us” private.

PAI runs entirely in RAM and never touches your computer’s hard drive. The host OS is never mounted. All AI inference happens locally via Ollama; your prompts never leave your machine. When you shut down, RAM is cleared and no forensic trace remains — no chat history, no downloaded files, no browser cookies, nothing.

The network stack adds further protection: your MAC address is randomized on every boot, the firewall denies all incoming connections, and Ollama binds only to localhost. Even if someone is on the same Wi-Fi network, they cannot reach your AI session.

For maximum privacy, enable Tor with sudo pai-privacy on to route all outbound traffic through the Tor network.

Tails is excellent for anonymous browsing and communication. PAI solves a different problem: private local AI.

	Tails	PAI
Primary purpose	Anonymous browsing	Private local AI
Tor integration	Always-on (everything)	Opt-in
AI capabilities	None	Full local LLM via Ollama
Desktop	GNOME	Sway (lighter)
Amnesic	Yes	Yes
ISO size	~1.9 GB	~912 MB

If your goal is anonymous web browsing and encrypted communication, use Tails. If your goal is conversing with an AI model without sending your prompts to a corporation, use PAI.

Many users who need both run PAI with Tor enabled — you get PAI’s local AI + Tails-style network anonymity in one system.

Everything in RAM is gone — permanently. This includes:

• All AI chat history and conversations
• Any files you downloaded or created
• Browser history, cookies, and saved passwords
• Wi-Fi credentials and connection history
• Shell history and command logs
• Every keystroke, clipboard entry, and screen capture

The host machine returns to exactly how it was before you booted PAI. There is no forensic residue.

This is called amnesic operation — the same design philosophy used by Tails. It means PAI is ephemeral by default: great for privacy, but you can’t resume where you left off without the opt-in encrypted persistence feature.

With encrypted persistence (opt-in via sudo pai-persistence setup), selected data is saved to a LUKS-encrypted partition on the USB drive and restored at the next boot after you enter your passphrase.

Yes — with opt-in encrypted persistence.

By default, PAI is amnesic: everything is wiped at shutdown. To save data across sessions, set up the encrypted persistence partition:

# First-time setup (creates LUKS partition on your USB)
sudo pai-persistence setup

# Save your data manually at any time
pai-save

# Data is also auto-saved on shutdown

What gets saved:

• Your home directory (~/)
• SSH keys (~/.ssh/)
• GPG keyring (~/.gnupg/)
• Electrum wallet (~/.electrum/)
• Downloaded AI models (so you don’t re-download each session)
• PAI UI chat history and settings

The persistence partition is encrypted with LUKS2 (AES-256). At boot, PAI asks for your passphrase on a secondary terminal (tty2). If you skip it, PAI boots normally in amnesic mode.

Your passphrase is the only key to your data. There is no recovery if you lose it.

No — PAI is designed to work fully offline.

The AI inference engine (Ollama) runs entirely on your hardware. Once you’ve downloaded a model, you can disconnect from the internet and keep chatting. Your prompts never leave your machine.

What needs internet (first time only):

• Downloading AI models via ollama pull <model> — models are 1–8 GB each
• Accessing web content in Firefox

What never needs internet:

• Running AI conversations with already-downloaded models
• Accessing the PAI chat interface at localhost:8080
• Everything else in PAI’s default workflow

With persistence enabled: download your models once, save them, and they’re available offline on every subsequent boot.

Air-gapped operation: if you need complete network isolation, PAI works perfectly with the Wi-Fi/Ethernet physically disconnected. The firewall also supports this: all AI services are localhost-only by default.

PAI ships without any pre-bundled models to keep the ISO small (~912 MB). You download models on first use.

Any model supported by Ollama works — that’s hundreds of models:

Model	Size	Best for
phi3:mini	~2 GB	Fast, low RAM, capable
llama3.2:3b	~2 GB	Fast, good quality
llama3.1:8b	~5 GB	High quality, balanced
mistral:7b	~4 GB	Fast, strong reasoning
gemma2:9b	~5 GB	Google’s model, great quality
codellama:7b	~4 GB	Code-focused tasks
deepseek-coder-v2	~9 GB	Advanced coding

To pull a model:

ollama pull llama3.2

Or through the PAI chat interface at localhost:8080.

Hardware guidance:

• 4 GB RAM: use phi3:mini or llama3.2:3b
• 8 GB RAM: use llama3.1:8b or mistral:7b
• 16 GB+ RAM: use larger 13B+ models

Pre-bundling roadmap: a future release will include a small model (e.g., Phi-3 Mini) baked into the ISO for instant offline use without any downloads.

PAI works on most x86_64 computers made in the last 15 years.

Confirmed to work:

• Any Intel or AMD 64-bit laptop or desktop
• Intel Macs (boot via Option key at startup)
• Most hardware from 2010 onwards

Not supported:

• Apple Silicon (M1/M2/M3/M4) — ARM64 build is on the roadmap
• 32-bit x86 processors
• Machines without USB boot capability (very old BIOS)

Minimum requirements:

• CPU: x86_64 (any speed — Ollama runs on slow CPUs, just slower)
• RAM: 2 GB (4 GB+ strongly recommended for AI inference)
• USB: ability to boot from USB (almost all modern machines support this)

BIOS/UEFI:

• Legacy BIOS: works
• UEFI: works (PAI uses a hybrid ISO that supports both)
• Secure Boot: must be disabled in UEFI settings

Common boot menu keys:

Manufacturer	Key
Dell	F12
HP	F9
Lenovo/ThinkPad	F12
ASUS	F8 or Esc
Acer	F12
Apple (Intel)	Option (⌥)
Generic UEFI	F11 or Del

If PAI boots but the screen is blank, try adding nomodeset to the GRUB kernel line.

Re-flash your USB with the latest ISO.

PAI doesn’t have an in-place upgrade mechanism — because it runs from a read-only squashfs, updates mean replacing the ISO on the USB drive.

Update process:

# Download the latest ISO
curl -LO https://get.pai.direct/pai-amd64.iso

# Verify checksum
curl -LO https://get.pai.direct/pai-amd64.iso.sha256
sha256sum -c pai.iso.sha256

# Re-flash (from a different machine while PAI USB is unplugged)
sudo dd if=pai.iso of=/dev/sdX bs=4M status=progress && sync

With persistence: re-flashing replaces the ISO partition but leaves your encrypted persistence partition intact. Your saved models, keys, and data survive the update.

Without persistence: since PAI is amnesic, there’s nothing to migrate — just boot the new version.

Staying informed:

• Watch GitHub Releases for new versions
• The Changelog app shows release notes
• Critical security releases will be announced prominently

Frequency: PAI is under active development. Major releases happen every few months; security patches are released as needed.

Three reasons: live-build, stability, and ecosystem.

live-build is Debian’s toolchain for building live ISOs. It’s battle-tested by Tails, Kali, and Parrot. Building a live USB OS that boots reliably across hundreds of hardware configurations is hard — live-build solves most of that problem. Porting to Arch or Alpine would require rebuilding that entire toolchain from scratch.

Stability. Debian Stable (Bookworm) has predictable behavior: packages don’t break between runs, the kernel is well-tested, and security patches come reliably for 5 years. Arch’s rolling release is great for personal machines but harder to pin for a reproducible OS build.

Ecosystem. Debian has 59,000+ packages in its repositories. Tails, Kali, and Parrot all chose Debian for the same reasons. The live USB community tooling, documentation, and institutional knowledge is deeply Debian-oriented.

Alpine’s tiny size is appealing but its musl libc introduces compatibility headaches with binaries like Ollama that are built against glibc.

Could PAI work on Arch or Alpine? Probably. Would it be worth the engineering cost compared to the user benefit? Not yet.

Yes — PAI’s system integration code is released under the GNU General Public License v3.0 (GPL-3.0).

You are free to:

• Use PAI for any purpose, including commercial
• Study how PAI works and modify it
• Redistribute PAI to others
• Distribute your modified versions

The only condition: if you distribute a modified version, you must make the source code available under the same GPL-3.0 license.

Included software retains its own licenses:

• Ollama: MIT license
• PAI UI (derived from LobeChat): MIT license
• Debian packages: various (GPL, LGPL, BSD, MIT) — all DFSG-compliant
• Firefox ESR: Mozilla Public License 2.0

In practice: you can fork PAI, add your own tools, and distribute it — as long as you publish your fork’s source code. The GPL ensures PAI remains open and improvements flow back to the community.

See the LICENSE file for the full text.