Nullcloud — On-Device, No Cloud AI for Any Page or PDF

AI that reads the web for you — without the web reading you. Strictly on-device, no cloud AI. Summarize, fact-extract, translate, or question any web page, text selection, or PDF using a local model running on your own phone or computer. Works completely offline in Airplane Mode.

---

Why Nullcloud Stands Out

1. Own the "Airplane Mode" Niche (On-Device, No Cloud AI): Unlike cloud wrappers, Nullcloud runs the LLM directly on your phone (127.0.0.1). Open an article, switch your phone to Airplane Mode, and stream a summary or ask questions with zero network connection. (We never call this "air-gapped" because your phone connects to the internet when you browse—it is strictly on-device, no cloud AI.)
2. Android Share Sheet -> Instant Summary Notification (No Extension Needed): In any Android browser or app, tap Share -> Termux, pick an action (Summarize, Key facts, Claims check, Read aloud), and receive a summary notification with Copy and Open UI buttons. No mobile browser extension required.
3. Unified One-Command Setup (bash setup.sh): A single setup command automatically detects Android Termux vs Desktop (macOS/Linux) and deploys all three entry points at once:
   • Android Share Sheet (~/bin/termux-url-opener)
   • Web UI & Productivity Hub (http://127.0.0.1:8080 — works in any browser)
   • Browser Extension (Chrome, Brave, Edge, Firefox)
4. GPU Offload on Android (Vulkan): termux/setup.sh checks /proc/meminfo, picks a default model by RAM (Llama 3.2 1B under ~5.5 GB, Llama 3.2 3B on 6 GB+, both Q4_K_M), or uses your own GGUF (e.g. Qwen3 4B) if you link it to ~/nullcloud-models/model.gguf, and offloads layers to the GPU (-ngl 99). Tested on a Samsung Exynos phone with GPU (Vulkan). Other phones may work but are not tested yet.
5. Everything Is Free: Library, PDF input, presets and Private Cloud are free for everyone. No license key, no paywall.

---

Quick Start — Unified Setup

Option A: Android Phone (Termux — Share Sheet + Web UI + Private Cloud)

Requirements: 4 GB+ RAM Android phone (~6 GB recommended for 3B model), ~2.5 GB free storage, Wi-Fi for the one-time model download.

1. Install Termux and Termux:API from F-Droid (not Play Store):
   • Termux on F-Droid
   • Termux:API on F-Droid
2. In Termux, run the unified setup command:

   pkg install -y git && git clone https://github.com/kukirin-ui/Nullcloud ~/nullcloud && bash ~/nullcloud/setup.sh

   (To force the fast 1B model on any phone: MODEL_TIER=1b bash ~/nullcloud/setup.sh; to compile llama.cpp from source with full mobile Vulkan GPU shaders: BUILD_VULKAN=1 bash ~/nullcloud/setup.sh)
3. Start Nullcloud:

   bash ~/nullcloud/termux/start.sh

4. Use any of the 3 entry points:
   • Android Share Sheet: In any app or browser, tap Share -> Termux -> pick Summarize -> get an on-device summary notification.
   • Web UI: Open http://127.0.0.1:8080 in any browser on your phone, paste your token once, and run.
   • Airplane Mode Test: Open an article or paste text/PDF into the Web UI, turn on Airplane Mode, and press Run On-Device.
5. Set Android Settings -> Apps -> Termux -> Battery -> Unrestricted so Android does not kill background generation.

Option B: Desktop (macOS / Linux / Windows)

1. Install Ollama and Python 3.
2. Run the unified setup & start script:
   • macOS / Linux: bash setup.sh (or bash desktop/start.sh)
   • Windows: desktop\start.bat
3. Load the Browser Extension:
   • Open chrome://extensions -> enable Developer mode -> Load unpacked -> select the extension/ folder.
   • Open any web page (even offline in Airplane Mode), click Nullcloud, paste your token once, and click Run on this page.

---

Performance

Tested on a Samsung Exynos phone (Xclipse 960 GPU) using GPU offload (Vulkan, packaged llama-cpp-backend-vulkan). Tune GPU offload layers (-ngl) in the Web UI (Private Cloud & Services -> Hardware Tweaks) or the browser extension.

Measured with llama-bench -p 512 -n 32 -r 1 (Vulkan, -ngl 99) inside the Meridian app Console on a real phone:

GPU	Model	Threads	Prompt (pp512)	Generation (tg32)
Samsung Xclipse 960 (Exynos)	Qwen3 4B Q4_K_M	2	232.29 t/s	23.23 t/s
Samsung Xclipse 960 (Exynos)	Qwen3 4B Q4_K_M	4	228.05 t/s	21.92 t/s
Samsung Xclipse 960 (Exynos)	Qwen3 4B Q4_K_M	8	167.78 t/s	13.79 t/s
Samsung Xclipse 960 (Exynos)	Llama 3.2 3B Q4_K_M	2	312.91 t/s	28.90 t/s

Real llama-bench run inside the Meridian app Console on Samsung Xclipse 960 (Vulkan)

Bonus: Llama 3.2 3B (-p 512 -n 128 -r 5, 2 threads). It is faster than Qwen3 4B because it is a smaller model (3.2B vs 4B parameters); Qwen3 4B stays the pick for smarter answers and code.

With full GPU offload, more CPU threads do not help: THREADS=2 is recommended (fastest prompt reading and generation). termux/start.sh uses 2 threads by default when GPU offload is on; override with THREADS=4 bash ~/nullcloud/termux/start.sh.

Measure your own device: run bash termux/bench.sh 20 in Airplane Mode with the charger unplugged. It prints real end-to-end tok/s and seconds per article. PRs with your results are welcome.

---

Everything Is Free

Nullcloud is 100% free and open source under the MIT License. Every feature is included for everyone:

Feature	Included
On-Device, No Cloud AI (Airplane Mode ready)	Yes
Android Share Sheet -> Summary Notification	Yes
Web UI (127.0.0.1:8080) & Browser Extension	Yes
7 Built-In Modes (Summary, Facts, Claims, ELI5, Translate, Ask, Recipe)	Yes
Read Aloud (TTS) & Markdown / Obsidian Export	Yes
Saved-Page Library with SQLite FTS5 Search	Yes
Offline PDF Input	Yes
Custom Prompt Presets (Legal, Journalism, Clinical + your own)	Yes
Private Cloud (phone as AI server for your laptop over Tailscale)	Yes

Services (Optional, Paid)

The app is free. If you want someone to set it up for you, these are paid services:

Service	Price	How to order
Remote setup (Termux, model, Share Sheet, Web UI on your phone or PC)	€30	Pay with Revolut
Setup + priority support (remote setup + 30 days of priority email help)	€45	Pay with Revolut
Private Cloud (Tailscale) setup (phone + laptop, tested end to end)	€60-80	Email Pozun.alen96@gmail.com for a quote
Team setup & consulting	from €250	Email Pozun.alen96@gmail.com for a quote

After paying, email Pozun.alen96@gmail.com with the payment confirmation to book a time.

Private Cloud: Use Your Phone as Your Laptop's AI Server (Free)

This uses Tailscale (free for personal use), a private encrypted network between your own devices.

1. Install Tailscale on your phone and laptop and log in to the same account on both.
2. On your Android phone in Termux:

   TAILSCALE=1 bash ~/nullcloud/termux/start.sh

   Nullcloud binds to your phone's 100.x.y.z Tailscale address. llama-server stays on 127.0.0.1:8081 and is never exposed.
3. On your laptop, open the Nullcloud browser extension -> Compute & Private Cloud Settings -> enter http://<phone-tailscale-ip>:8080 and your token. Or open that address in a browser.

Note: if tailscale is not available inside Termux, Nullcloud keeps listening on 127.0.0.1 only. Set HOST=<phone-tailscale-ip> manually in that case.

High-Compliance Consulting (Law Firms, Journalists, Medical Clinics)

Need turn-key deployment of on-device, no cloud AI for privileged legal discovery, investigative source protection, or clinical literature workflows?

• Read docs/CONSULTING.md for vertical presets, threat models, and deployment packages.
• See Services above for prices.

---

How Nullcloud Compares

Capability	Nullcloud	Page Assist	Cloud AI Extensions
Branding & Architecture	On-device, no cloud AI	Local / Ollama	Cloud APIs
Android Share Sheet -> Summary Notification	Yes (No extension needed)	Not offered	Requires app/cloud
Runs 100% Offline in Airplane Mode on Phone	Yes (Termux + 1B/3B GGUF)	Desktop-focused	No
Phone as Private Cloud Server over Tailscale	Yes (free)	Manual	N/A
Offline Saved-Page FTS5 Library & PDF Input	Yes (free)	Chat history	Cloud-stored
Markdown / Obsidian Vault Auto-Export	Yes (NOTES_DIR)	No	No

---

Troubleshooting (Termux)

• Install Termux from F-Droid, not the Play Store. The Play Store build is outdated and pkg fails.
• pkg install prints Abort. The prompt was cancelled (no input or a stray key). Run it again with -y, for example pkg install -y nodejs git python.
• Mirror / 404 errors during pkg install: run termux-change-repo, pick a different mirror, then pkg update -y && pkg upgrade -y.
• No space left on device: the 3B model needs ~2 GB plus ~0.5 GB for tools. Free space or force the 1B model: MODEL_TIER=1b bash ~/nullcloud/setup.sh.
• Model download stopped halfway: run setup again. Partial *.gguf.part files are resumed or replaced.
• Share -> Termux does nothing: install Termux:API from F-Droid as well, run pkg install -y termux-api, and check that ~/bin/termux-url-opener exists and is executable (chmod +x ~/bin/termux-url-opener).
• Generation stops when the screen turns off: Android Settings -> Apps -> Termux -> Battery -> Unrestricted, and run termux-wake-lock.
• GPU is not used (llama-bench shows backend CPU, or start.sh says "CPU fallback"): install the Vulkan backend and the Android Vulkan loader:

  pkg install -y llama-cpp llama-cpp-backend-vulkan
  pkg uninstall -y vulkan-loader-generic
  pkg install -y vulkan-loader-android
  llama-server --list-devices

  You should see a line like Vulkan0: Samsung Xclipse 960 (...). If the package is not available, build from source: BUILD_VULKAN=1 bash ~/nullcloud/setup.sh.
• llama-server fails with a --flash-attn error: newer llama.cpp needs a value (--flash-attn on). Update Nullcloud with cd ~/nullcloud && git pull; start.sh also falls back to GPU without Flash-Attention.
• Web UI says "Wrong token": your token is in ~/.nullcloud-token. Show it with cat ~/.nullcloud-token and paste it once in the Web UI.
• Port 8080 already in use: bash ~/nullcloud/termux/stop.sh, then start again.

Still stuck? Open an issue, or book a remote setup.

---

Updating & GitHub tokens

This repository is public. You do not need a GitHub token or login to git clone or git pull.

Update Nullcloud on your phone (Termux):

cd ~/Nullcloud && git pull && pkill -f server.py; bash termux/start.sh

Then refresh http://127.0.0.1:8080.

If git asks for a username or password, your remote URL is wrong. Fix it:

git remote set-url origin https://github.com/kukirin-ui/Nullcloud.git

Then run git pull again. Do not paste a GitHub token as the password for a public clone/pull.

A GitHub token (fine-grained PAT) is only needed if you want to git push changes. Never paste a GitHub token into the Nullcloud app, never commit it, and never share it. (The separate Nullcloud Web UI token in ~/.nullcloud-token is unrelated — that is only for the local API.)

Recipe preset: paste a page URL, choose Balanced (a few relevant links) or Max (deeper crawl), extract facts into a saved recipe under ~/nullcloud-recipes/, and stop anytime — partial results still save. Open them from the Library tab (Saved Recipes).

Configuration & Environment Variables

Variable	Default	Description
TOKEN	~/.nullcloud-token	Required secret token sent in X-Token header
HOST	127.0.0.1	Bind address (TAILSCALE=1 auto-selects your 100.x.y.z Tailscale IP)
PORT	8080	Web UI and HTTP API port
LLM_URL	http://127.0.0.1:8081/v1/chat/completions	Local OpenAI-compatible completion endpoint (llama-server or Ollama)
MODEL_TIER	Auto (1b or 3b)	Force 1b (4 GB RAM) or 3b (6 GB+ RAM) in termux/setup.sh
GPU_LAYERS	99	Number of transformer layers offloaded to GPU (-ngl / num_gpu)
CPU_THREADS	Auto (4 on 1B, 6 on 3B)	Number of CPU performance threads used for prompt evaluation & generation
BUILD_VULKAN	0	Set BUILD_VULKAN=1 during termux/setup.sh to compile llama-server with mobile Vulkan GPU shaders
MAX_CHARS	6000 (4500 on 1B)	Maximum extracted page/PDF characters sent to local model
MAX_TOKENS	400 (300 on 1B)	Maximum generated completion tokens
NOTES_DIR	(disabled)	Opt-in path to save every output as .md (e.g. Obsidian vault)
ARCHIVE	0	Set ARCHIVE=1 with NOTES_DIR to cache raw HTML for offline re-reading

---

Support the Project

• Donate (optional): Revolut. Any amount helps.
• Star the repo and share your bench.sh results.
• Sibling Project: Check out Meridian — the open-source Android command deck and local AI mesh for Termux.

License

MIT License © 2026 Alen Požun (kukirin-ui)