Private local model · No subscription · Yours for life

Your AI.
Your machine.
No one else's business.

A.B.S.O.L is a private, local AI — the model lives on your hardware, not in someone's cloud. Chat, agent tools, voice, and image generation, with a live privacy console that proves nothing leaves your computer. Buy it once and it's yours for life — no subscription, no per-token billing, no account.

Pay once, own forever — runs offline. macOS / Linux: python3 app/install.py

0
cloud calls by default
100%
of inference on your GPU
1 file
auditable privacy log
$0/mo
pay once — yours for life

What's inside

A full assistant, not a chat box.

Everything below runs on your machine — see Editions for what each tier includes. The only two optional cloud features are individually flagged — and hard-blocked under Privacy Mode.

🧠

Local language model

Runs Hermes-3 Llama-3.1 8B (or any .gguf model you choose) through KoboldCpp on your own GPU. No API keys, no per-token billing, no outages.

🛠️

Agent tools

ABSOL can run sandboxed code, read and write files inside a contained workspace, and search the web — with every capability gated behind the privacy allowlist.

🛡️

Privacy Fortress

A deny-by-default network allowlist wraps every outbound request. Flip PRIVACY_MODE and only localhost traffic survives — enforced, logged, and visible live in the app.

🔮

Living avatar

The Metatron's Cube emotional avatar reacts to what ABSOL is actually doing — seven emotion states drive its color, flicker, and animation in real time.

🎙️

Voice, fully local

Two-stage local text-to-speech (Piper + RVC voice conversion) gives ABSOL a real voice — synthesized entirely on your machine, with mouth-sync driving the avatar.

🎨

Local media generation

Generate images with Stable Diffusion and video with LTX-Video, on your own GPU, behind a mandatory content-safety filter. Nothing is uploaded anywhere.

Privacy, enforced — not promised

Trust the log, not the marketing.

Most "private" AI asks you to take its word for it. ABSOL shows you the enforcement, live, and writes every network decision to a log you can read yourself.

  • Deny-by-default allowlistEvery outbound request is checked against an explicit host list. Lookalike domains are tested and defeated.
  • Hard Privacy ModeOne switch blocks all non-localhost traffic outright — including web search. Blocked means blocked, not "warned".
  • OS-level firewall lockdownAn included script blocks the inference engine from ever reaching the internet at the Windows Firewall level.
  • Contained file accessAgent file tools are locked to a workspace directory — path traversal and absolute-path escapes are tested against directly.
privacy_log.jsonl — live
12:04:31 ALLOW localhost:5001 kobold inference 12:04:33 ALLOW localhost:7860 image generation 12:04:40 BLOCK telemetry.vendor.com not on allowlist 12:05:02 PRIVACY_MODE=1 engaged 12:05:09 BLOCK html.duckduckgo.com cloud denied 12:05:09 ALLOW localhost:5001 kobold inference 12:05:17 ALLOW localhost:8420 absol console

Sentinel edition

Four watchers for when the shield comes down.

Privacy Mode is absolute — but sometimes you want ABSOL to reach the internet. The moment you allow it, the Sentinel Array activates: four autonomous defensive agents, each watching a different security surface, reporting live to the privacy console. Re-engage local-only mode and they stand down.

VANGUARDACTIVE — clear

Network watch

Watches live traffic for anomalies — burst activity to a single host, unexpected destinations, volume spikes that look like exfiltration.

AEGISACTIVE — scanning

Exfiltration shield

Scans every AI response before it reaches you, flagging SSNs, card numbers, file paths, and other PII that has no business appearing. Stores only pattern type — never the content itself.

CIPHERACTIVE — correlating

Injection sentry

Correlates prompt-injection attempts across conversations — catching the slow, distributed manipulation a single-message filter would miss.

ORACLEACTIVE — verified

Audit integrity

Continuously verifies the privacy log's hash chain, so tampering with the audit trail is detectable — the log you trust stays trustworthy.

Editions

Pay once. Private for life.

There is no tier that "unlocks" your own privacy — enforcement is identical everywhere. Higher editions add capability, never surveillance.

Core

$79 one-time

The complete private local AI.

  • Full local chat + agent tools on your GPU
  • Privacy Fortress & live console
  • Metatron's Cube emotional avatar
  • Local voice + image & video generation
  • One-shot Windows installer
Buy Core — $79

Workshop

$199 one-time

For tinkerers who want to make ABSOL theirs.

  • Everything in Sentinel
  • LoRA fine-tuning kit + starter dataset
  • Custom voice training walkthrough (RVC)
  • Avatar & persona customization guide
  • Early builds
Buy Workshop — $199

All editions run 100% offline. Pay once, own it forever — no subscription, no account, no per-token billing.
A typical cloud AI subscription costs $240+ every year. ABSOL never bills you again.

Requirements

Runs on real hardware, not a data center.

Built and tested around a single mid-range GPU. If your machine can game, it can run ABSOL.

OSWindows 10/11 (INSTALL.bat)  ·  macOS / Linux (install.py)
GPUNVIDIA, 8 GB VRAM minimum — 12 GB recommended (RTX 3060-class) for voice + media generation
RAM16 GB recommended
Disk~10 GB for the app + your chosen model file
ModelAny .gguf — Hermes-3 Llama-3.1 8B recommended; download links included in the app
InternetOnly for the initial download. After that: optional, and blockable.

FAQ

Fair questions, straight answers.

Is it actually private, or "private" with an asterisk?

Actually private — and verifiable. All inference, voice, and media generation run on your GPU. Exactly two optional features touch the cloud (Wispr Flow speech-to-text and a stock-upload helper), both clearly flagged in the app, and both hard-blocked when Privacy Mode is on. Every network decision is written to a plain-text log you can open yourself.

Do I need an account or an API key?

No account, no API key, no sign-in, no telemetry. Download it, run the installer, drop in a model file, done.

What does it cost to run?

Electricity. There's no per-message or per-token billing because there's no cloud service behind it — the model is a file on your disk.

Can I use a different model?

Yes. ABSOL runs any .gguf model through KoboldCpp. It ships tuned around Hermes-3 Llama-3.1 8B, and the Workshop edition includes a full LoRA fine-tuning kit to train the persona onto your own model.

What are Sentinels?

Four autonomous defensive agents — VANGUARD, AEGIS, CIPHER, and ORACLE — that activate the moment you let ABSOL reach the internet. They watch network traffic for anomalies, scan every response for PII before it reaches you, correlate prompt-injection attempts across conversations, and verify the audit log hasn't been tampered with. They never store the sensitive content they flag, only metadata, and they stand down cleanly when you re-engage local-only mode. Core keeps you private by staying closed; Sentinel keeps you guarded when you choose to open the door.

What's with the avatar?

ABSOL renders as a Metatron's Cube — the sacred-geometry figure — with a fox companion badge. It isn't decorative idle animation: its color, flicker, and motion are driven by the agent's real state and a seven-emotion classifier, so you can see at a glance whether it's idle, working, or raising an alert.

Windows says the installer is unrecognized.

Expected for a small independent app without a code-signing certificate. The package is plain batch + Python you can read before running — nothing is obfuscated.

Bring your AI home.

Two steps on Windows: run INSTALL.bat, then click the fox on your desktop.

Buy Sentinel — $149 Buy Core — $79 Compare editions

Secure checkout · Your download link and license are delivered instantly after payment.