PortableAI — Plug-and-play Local LLM Server

Run any GGUF language model locally — no Python, no Docker, no cloud, no install headaches.
Works from a USB drive. Runs on Linux, macOS, and Windows. One script to set up everything.

---

What is this?

PortableAI wraps llama.cpp's llama-server into a zero-dependency portable package. You plug in a USB drive (or clone the repo), drop in a model, and run a single script. A local web UI opens in your browser and you can start chatting — fully offline, fully private.

No Python environment. No package managers. No GPU required.

---

Features

• Truly portable — runs from a USB stick on any machine
• Cross-platform — Linux (x64 + arm64), macOS (Intel + Apple Silicon), Windows (x64)
• Auto-installer — fetches the correct llama-server binary for every platform in one run
• Model picker — prompts you to choose when multiple .gguf models are present
• CPU-only — works on any modern machine without a GPU
• 100% offline after setup — no data leaves your machine
• Built-in web UI — chat interface served directly by llama-server
• LAN sharing — accessible from any device on the same Wi-Fi at http://0.0.0.0:8080

---

Directory Layout

PortableAI/
├── install.sh          ← Linux/macOS installer (all platforms)
├── install.bat         ← Windows installer
├── start.sh            ← Linux/macOS launcher
├── start.bat           ← Windows launcher
├── models/             ← Drop your .gguf model files here
├── ui/                 ← (optional) custom web UI override
└── bin/
    ├── linux/
    │   ├── linux_x64/
    │   │   ├── llama-server-linux-x64   ← renamed server binary
    │   │   ├── libllama.so              ← required shared libs
    │   │   ├── libggml.so
    │   │   └── libggml-cpu.so  ...
    │   └── linux_arm64/
    │       └── llama-server-linux-arm   + .so libs
    ├── mac/
    │   ├── mac_arm64/
    │   │   └── llama-server-mac-arm     + .dylib libs
    │   └── mac_x64/
    │       └── llama-server-mac-x64     + .dylib libs
    └── windows/
        ├── llama-server-win.exe
        └── *.dll                        ← required DLLs

---

Quick Start

Step 1 — Get a model

Download any .gguf model and place it in the models/ folder.

Recommended for most machines (4–8 GB RAM):

Search huggingface.co for any model — filter by GGUF format and pick a Q4_K_M quantization for the best balance of size and quality.

Step 2 — Install binaries

Linux / macOS

chmod +x install.sh
./install.sh

The installer asks which platform(s) to set up:

 Select which platform(s) to install:

   [1] Linux x64 (most PCs/servers)
   [2] Linux arm64 (Raspberry Pi, ARM servers)
   [3] macOS arm64 (Apple Silicon M1/M2/M3)
   [4] macOS x64 (Intel Mac)
   [5] Windows x64 (CPU)

   [A] All platforms (for a shared USB drive)
   [Q] Quit

 Tip: enter multiple numbers separated by spaces (e.g. 1 3)

Choose A to pre-load all platforms if you're preparing a USB drive that will be used on multiple machines.

Windows

Double-click install.bat

Step 3 — Start the server

Linux / macOS

chmod +x start.sh
./start.sh

Windows

Double-click start.bat

The browser opens automatically at http://127.0.0.1:8080

If you have multiple models, you'll be asked to choose one:

 [?] Multiple models found — select one:
     ─────────────────────────────────────────────
     [1] mistral-7b-instruct-v0.2.Q4_K_M.gguf     3.8G
     [2] llama-3.2-3b-instruct-q4_k_m.gguf         1.9G

 Enter number [1-2]:

---

Requirements

Runtime

Platform	Requirement
Linux	glibc 2.17+ (any modern distro)
macOS	11.0+ (Big Sur or later)
Windows	Windows 10 (build 17063+), Visual C++ Redistributable

For install.sh

Tool	Notes
curl	Pre-installed on most systems
tar	Pre-installed on all Linux/macOS
unzip	Only needed for Windows zip on Linux: sudo pacman -S unzip

Hardware (minimum)

RAM	Recommended model size
4 GB	1B–3B models (Q4_K_M)
8 GB	7B models (Q4_K_M)
16 GB	13B models (Q4_K_M)
32 GB	30B+ models (Q4_K_M)

---

How It Works

install.sh / install.bat
        │
        ├── Queries GitHub API for latest llama.cpp release
        ├── Downloads the correct archive per platform:
        │       Linux x64   → llama-bXXXX-bin-ubuntu-x64.tar.gz
        │       Linux arm64 → llama-bXXXX-bin-ubuntu-arm64.tar.gz
        │       macOS arm64 → llama-bXXXX-bin-macos-arm64.tar.gz
        │       macOS x64   → llama-bXXXX-bin-macos-x64.tar.gz
        │       Windows x64 → llama-bXXXX-bin-win-cpu-x64.zip
        ├── Extracts ALL files (binary + shared libs / DLLs)
        └── Renames llama-server → platform-specific canonical name

start.sh / start.bat
        │
        ├── Scans models/ for .gguf files
        ├── Prompts for selection if > 1 model found
        ├── Detects OS + architecture
        ├── Sets LD_LIBRARY_PATH (Linux) / DYLD_LIBRARY_PATH (macOS)
        │   so shared libs next to the binary are always found
        └── Launches llama-server on port 8080

Why copy all the .so / .dll files?
llama-server dynamically links against libllama, libggml, libggml-cpu and others. Copying only the executable would cause an immediate crash with a "shared library not found" error. The installer copies the entire archive contents into the bin directory, and start.sh sets LD_LIBRARY_PATH to that directory so the binary is fully self-contained without touching any system paths.

---

Configuration

The server starts with sensible defaults. To customize, edit the exec "$BIN" block at the bottom of start.sh / start.bat:

exec "$BIN" \
    -m "$MODEL"      \   # model file path (set automatically)
    -c 4096          \   # context window size (tokens)
    -t "$THREADS"    \   # CPU threads (auto: nproc - 1)
    --port 8080      \   # HTTP port
    --host 0.0.0.0       # bind address (0.0.0.0 = LAN accessible)

Common tweaks:

Flag	Example	Effect
-c	-c 8192	Larger context (needs more RAM)
-t	-t 4	Fixed thread count
--port	--port 9090	Change the port
--host	--host 127.0.0.1	Localhost only (disable LAN)
-ngl	-ngl 35	Offload layers to GPU (if available)

Full flag reference: ./bin/linux/linux_x64/llama-server-linux-x64 --help

---

Accessing from Other Devices

While the server is running, any device on the same network can access the UI:

1. Find your machine's local IP: ip addr (Linux) / ipconfig (Windows)
2. Open http://192.168.x.x:8080 on the other device

---

Troubleshooting

llama-server: error while loading shared libraries: libllama.so
The shared libs are missing from the bin directory. Re-run install.sh — it copies all files from the archive, not just the binary.

Binary not found / install.sh stops silently after printing the release tag
GitHub API rate-limited the request (60 requests/hour for unauthenticated IPs). Wait a few minutes and try again.

VCRUNTIME140_1.dll not found (Windows)
Install the Visual C++ Redistributable and re-run start.bat.

Model loads but responses are very slow
Use a smaller or more quantized model (e.g., Q2_K instead of Q8_0). Reduce context with -c 2048.

Port 8080 already in use
Change --port 8080 to another port (e.g. --port 9090) in start.sh / start.bat.

---

Updating llama.cpp

Just re-run the installer. It always fetches the latest release from GitHub:

./install.sh   # picks what to update, overwrites existing binaries

---

Project Structure

Portable_Local_AI/
├── install.sh      Universal installer — Linux/macOS, all platforms
├── install.bat     Windows installer
├── start.sh        Launcher — Linux/macOS
├── start.bat       Launcher — Windows
├── models/         Your .gguf model files go here
├── ui/             Optional: override the built-in web UI
└── bin/            Auto-populated by installer

---

Credits

• llama.cpp by ggml-org — the inference engine powering everything
• Models from HuggingFace — community-converted GGUF weights

---