Plug in. Run one script. Get a fully autonomous AI agent with 13 tools, deep search, OSINT, and persistent memory — all from a portable USB/SSD. Zero install, zero traces.
The original USB-Uncensored-LLM (now archived) was a great portable chat. This takes it further — it's not just a chat, it's an agent that can actually DO things.
| Feature | Original | This Project |
|---|---|---|
| Chat | ✅ | ✅ |
| Web Search | ❌ (broken) | ✅ DuckDuckGo Lite |
| Deep Search (multi-source) | ❌ | ✅ Fetches + reads + synthesizes |
| OSINT (email lookup) | ❌ | ✅ Holehe (121+ sites) |
| OSINT (username search) | ❌ | ✅ Maigret (600+ platforms) |
| OSINT (social search) | ❌ | ✅ Agent Reach (15 platforms) |
| File Operations | ❌ | ✅ Read/Write/List |
| Code Execution | ❌ | ✅ Python + Shell |
| Persistent Memory | ❌ | ✅ Auto-saves to USB |
| Cross-Platform | Partial | ✅ Windows + Linux |
| Autonomous Tool Calling | ❌ | ✅ Model decides which tools to use |
# First time: install engine + models (interactive menu)
Windows\install.bat
# Unattended variants
Windows\install.bat -Models 1 -Yes # engine + Gemma 2B only, no prompts
Windows\install.bat -Models 1,3 # two models
Windows\install.bat -SkipModels # engine + Python only, models later
# Every time: launch the agent
Windows\start-agent.batinstall.bat calls Windows\install-core.ps1 (same folder) — that script does all the
work: portable Python, portable Ollama engine, GGUF downloads, Modelfiles, import.
# First time: install engine + models
bash Linux/install.sh
# Every time: launch the agent
bash Linux/start-agent.shA local web chat UI ships alongside the CLI (Linux/start.sh / Windows/start-fast-chat.bat) — no cloud, no account:
The agent has 13 tools it can use autonomously:
| Category | Tool | Description |
|---|---|---|
| 🔍 Search | web_search |
Search the web (DuckDuckGo Lite) |
| 🔍 Search | deep_search |
Multi-source research (search → fetch → synthesize) |
| 🔍 Search | web_get |
Download and extract any webpage |
| 🕵️ OSINT | osint_email |
Check where an email is registered (121+ sites) |
| 🕵️ OSINT | osint_username |
Find a username across 600+ platforms |
| 🕵️ OSINT | osint_social |
Search 15 social platforms simultaneously |
| 📁 Files | read_file |
Read any file on the system |
| 📁 Files | write_file |
Create or overwrite files |
| 📁 Files | list_files |
List directory contents |
| ⚡ Exec | run_command |
Execute shell commands |
| 🐍 Exec | run_python |
Execute Python code |
| 🧠 Memory | remember |
Save information persistently |
| 🧠 Memory | recall |
Retrieve saved information |
Pre-configured for 5 uncensored GGUF models (~25GB total):
- Gemma 2 2B Abliterated (~1.7 GB) — Blazing fast, recommended for all
- Gemma 4 E4B Ultra Uncensored (~5.3 GB) — Maximum compliance
- Qwen 3.5 9B Uncensored (~5.6 GB) — Best for agent/tool calling
- NemoMix Unleashed 12B (~7.5 GB) — Heavyweight reasoning
- Dolphin 2.9 Llama 3 8B (~4.9 GB) — Classic uncensored
Models are downloaded once and work on any OS without re-downloading.
- USB 3.0 drive or SSD with 30+ GB free
- Host PC: 8 GB RAM minimum (16 GB recommended for 9B/12B models)
- Linux:
python3(built-in on most distros) - Windows: nothing (auto-downloads portable Python on first run)
USB-AI-Agent/
├── Windows/
│ ├── install.bat # Entry point: downloads Python + Ollama + models
│ ├── install-core.ps1 # The actual installer (called by install.bat)
│ ├── start-agent.bat # 🆕 Launches the autonomous agent
│ └── start-fast-chat.bat # Original web chat UI
├── Linux/
│ ├── install.sh # Downloads Ollama Linux + models
│ ├── start-agent.sh # 🆕 Launches the agent
│ └── start.sh # Original web chat
├── Shared/
│ ├── agent.py # 🆕 The autonomous agent (13 tools)
│ ├── chat_server.py # Web chat server
│ ├── FastChatUI.html # The web chat interface it serves
│ ├── scripts/ # Shared model catalog + UI asset downloader
│ ├── python/ # Auto-downloaded portable Python
│ ├── bin/ # Portable Ollama engine
│ ├── logs/ # Install logs
│ └── models/ # GGUF model files + Modelfiles
└── .gitignore
- Everything runs fully offline after initial download
- Zero data leaves your machine (except web searches when online)
- No accounts, no telemetry, no tracking
- Chat history saves only to your USB drive
- Works on Tails and other privacy-focused Linux distros
This repo also includes standalone Android applications built with Capacitor + React:
| App | File | Description |
|---|---|---|
| 🛸 Blu | android-apks/Blu-v1.0.apk |
AI talking friend — your personal AI companion |
| 🤖 ZENO | android-apks/ZENO-v1.3-fix.apk |
AI friend v1.3 — improved AI companion with fixes |
android-apks/AI-Flight-Search.apk |
AI-powered flight search app |
All apps are Capacitor-based (React + TypeScript) and can be side-loaded on any Android device.
| Symptom | What to do |
|---|---|
install.bat says install-core.ps1 is missing |
Fixed in this repo — that file used to be absent, so the Windows installer could not run at all. Update your copy of Windows\ (or re-clone) and run install.bat again. It is not your antivirus. |
| A download stops halfway | Just run install.bat again. Partial files resume (curl -C -) and finished ones are skipped. |
start-agent.bat says Ollama Engine Not Found |
The engine step did not complete. Re-run install.bat (it retries only what failed). |
| Import fails during setup | See the log written to Shared\logs\install-<timestamp>.log, then re-run install.bat. |
| PowerShell blocks the script | install.bat already passes -ExecutionPolicy Bypass; run it from cmd with Windows\install.bat -SkipModels to isolate the failing step. |
Requirements on the host PC: Windows 10/11 with curl.exe (built in since 1803) and
about 4 GB free beyond the models (portable Python ~11 MB, engine ~2 GB after extraction).
Built on top of USB-Uncensored-LLM by Tech Jarves.
OSINT tools: Holehe, Maigret, Agent Reach.
| Project | Description |
|---|---|
| 🛡️ Pusu AdBlocker | Chrome MV3 ad blocker — 35,000+ rules, zero tracking |
| 🌐 AI-Flight-Search | AI-powered flight search engine |
| 💡 ZENO | AI project — details coming soon |
MIT — same as the original project.