🔒 usb.plan.ai: Portable AI Council

Bring the best models you trust into one auditable council. Runs entirely from a USB drive. Offline-first · private · open-source. A deliberated answer with dissent shown, not one chatbot's opinion.

Inspired by PortableMind and the OSS DIY build techjarves/Portable-AI-USB. Its core is different: an auditable AI Council instead of a single model. Offline by default; online frontier council on opt-in.

Status: v0.3 concept & spec, building in public

Today this repo is the website, documentation, and v0.3 spec. The product is not built yet; see the status, roadmap, and contribute to help build it.

---

Why a council?

A single model gives you one answer you have to trust blindly. usb.plan.ai runs a 3-stage council (First Opinions → Peer Review → Chairman) and shows you the full trace: every opinion, the ranking, and where the models disagreed.

"Run this risky bash script offline."

  Solver     → just sudo it.
  Skeptic    → ⚠ sudo + unverified script = full-system exposure.
  Security   → no checksum, no sandbox. block.
  Chairman   → dry-run in a sandbox, verify the checksum, run as your user.

The Skeptic catches what a single chat misses. That's the whole point.

---

Council seats

Swappable "seats", not "the smartest models". The roster lives in Council seats; redistribution policy lives in Model licenses. The default bundle stays MIT/Apache only, with Gemma opt-in because its notice must travel with redistribution.

---

Setup

Requirements

Use a fast USB drive and run preflight first; tier sizing, performance, and filesystem rules live in hardware tiers, performance, and directory structure.

Steps

1. Download the latest release and unzip it onto the drive.
2. Run preflight: double-click preflight-check.*. It measures filesystem, USB read speed, RAM, and GPU/VRAM, then recommends a tier. (Run this first.)
3. Install: run install.*, pick a tier (Pocket / Field / Lab). Models download into model-packs/ with SHA-256 verification.
4. Done. Your portable council is ready.

---

Run it

Double-click the launcher for your OS. It serves the UI in your browser at http://localhost:4321.

OS	Launcher
Windows	start-windows.bat
macOS	start-mac.command
Linux	start-linux.sh

First run on a new machine can take a moment while the engine extracts once and is auto-cleaned on eject. Ask the demo prompt and watch the council trace.

---

Privacy

• Offline by default: nothing leaves the stick. UI shows offline.
• Opt-in online: escalate to the online frontier council only after a Privacy-Diff shows you exactly what's sent, to whom, and the cost. UI shows online-council.
• Vault: keys, chats, and traces are encrypted (Argon2id + AES-256), unlocked by your passphrase. Key material is never written plaintext.
• Minimal-trace, honestly: no telemetry, no accounts. But a portable app can't leave zero trace on every OS, and a compromised host is out of scope. We document the unavoidable artifacts instead of pretending they don't exist.

---

What's on the stick

usb.plan.ai/
├── start-windows.bat / start-mac.command / start-linux.sh
├── preflight-check.*      ← FS / USB-speed / RAM / GPU check (run first)
├── install.*              ← tier + model selection
├── engine/                ← llama.cpp llama-server, whisper.cpp, vision
├── core/                  ← FastAPI council orchestrator + council.config.yaml
├── ui/                    ← browser-only web UI
├── model-packs/           ← verified manifests (license + SHA-256)
├── recipes/ playbooks/    ← roles, seats, offline workflows
├── evals/                 ← Council-vs-Single proof data
├── plugins/               ← new providers / models / tools
└── vault/                 ← Argon2id + AES-256 (keys, chats, traces)

---

Hardware sizing

Hardware sizing lives in hardware tiers. Below Lab, use the local role council; Lab is the local multi-model council tier.

---

Important notes

• Code license ≠ model license. This repo is Apache-2.0; each model keeps its own weights license. See Model licenses and model-packs/ manifests.
• pnpm only for the docs/site build: npm/yarn silently break it.
• Never "just sudo" a downloaded script. Ask the council first.

---

Docs & contributing

Full documentation: usb.plan.ai/docs. It covers architecture, the council, models, hardware, security, config, evals.

It's a council kit, not a monolith. Add a recipe/, a model-pack/, a plugin/, or a playbook/. Good first issues are tagged. See docs → Contributing.

---

License

Apache-2.0 (code). Model weights are governed by their own licenses. See LICENSES.md.

_{Open. Offline. Auditable. Smarter together. plan.ai}