Welcome

usb.plan.ai is an open-source, portable AI stick with a fundamentally different core from existing offline-AI sticks: an auditable AI Council.

The core: instead of a single chat answer, usb.plan.ai gives you a deliberated, multi-perspective answer. Locally, that is a local role council (one strong model in several roles), and, when a network is available and you explicitly allow it, as a real multi-provider online frontier council. The Council is a deliberate review mode, not the default chat.

Status: concept / spec v0.3

The website and these docs exist today. The product is not built yet. That includes the portable engine, the council orchestrator, the evals, and the vault. This is a build-in-public spec; everything described here is the target. See the roadmap.

Positioning

A good single model wins at “write me an email.” A council wins at “review this decision / this code / this risk from several angles and show me the dissent.”

It is built for situations where internet access, privacy, or model dependence are problematic; see what makes it different and the Council.

Guiding principles

Principle	Meaning
Local-first, not local-only	Works offline. With a network it gets smarter, but never without explicit consent.
Minimal-trace (honest, not “zero”)	No persistent user data on the host. Unavoidable OS artifacts are documented openly.
Council > single model, but proven	Multi-perspective synthesis must be measurably better via evals, not asserted.
Open, forkable, extensible	A council kit with recipes, model-packs, and plugins. Not a monolith.
Bring-your-own-keys	Keys belong to the user, stored passphrase-encrypted on the stick.
Honest hardware expectations	Clear minimums per tier; no exaggerated plug-and-run claims.

At a glance

Domain	usb.plan.ai
Product	Portable AI Council
Tagline	”Bring the best models you trust into one auditable council. Offline-first.”
Code license	Apache-2.0 (model weights keep their own; see licenses)
Status	Concept / spec v0.3 (council-optimized), 2026-06-03
Owner	plan.ai

v0.3 is the result of a 3-stage LLM council (first opinions on Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro → peer review → Chairman synthesis). The key corrections versus v0.1 are in the council changelog.