NativeLab

A fully local, privacy-first LLM workbench powered by llama.cpp - desktop GUI, terminal CLI, and an experimentation layer.

NativeLab is a desktop and terminal client for running large language models entirely on your machine. It wraps llama.cpp, running Ollama models, optional Hugging Face Transformers models, and API backends behind a polished PyQt6 GUI and a Claude-Code-style terminal CLI, with first-class support for multi-model pipelines, document references, long-document summarization, and a brand-new Labs experimentation layer.

PhonoLab is the official Android client - same local-first philosophy, runs llama.cpp on-device via JNI, with a chat UI, document attachments, RAG, vision model support, and a LAN API server. See PhonoLab/ or the PhonoLab page.

pip install nativelab
nativelab            # GUI
nativelab --cli      # terminal control center (setup, chat, models, labs, integrations)

# Standalone local server (turn any GGUF into an API)
python -m nativelab.server --model models/llama-7b.gguf --port 8787 --host 0.0.0.0

App screenshots

‹

›

‹

›

‹

›

‹

›

‹

›

‹

›

‹

›

‹

›

✨ Highlights

• 🖥️ Desktop GUI - Chat, model library, visual pipeline builder, MCP, Download tab, Labs, theming.
• ⌨️ Terminal CLI - nativelab --cli opens a full terminal control center for chat, local/API models, skills, Labs, saved pipelines, integrations, endpoint serving, @file embedding, slash commands, and linting.
• 🧪 Labs - A dedicated experimentation layer with a shared endpoint API. New lab features get engine status, model swap, context change, and LLM calls for free.
• 🔌 Integrations - Local JSON endpoint, route browser, and saved Discord/WhatsApp bot connector profiles.
• 🔗 Visual Pipeline Builder - 20+ node types, shipped example presets, resizable sidebars, AI-assisted pipeline generation, native-accelerated graph helpers, live execution log, save/load.
• 🌐 Backend mixing - Local GGUF, running Ollama models, optional Hugging Face Transformers models, OpenAI-compatible APIs, and Anthropic endpoints share the same app state.
• 📱 LAN Device Discovery - Scan your local network for PhonoLab Android devices, register them as API model endpoints, route inference to phones and tablets over WiFi.
• 🔐 Hugging Face login - One-click browser login for gated/private repos, with access-token paste as an advanced fallback.
• ⚡ Parallel + pipeline mode - Run reasoning + coding engines simultaneously and chain them automatically.
• 🧠 Auto family detection - 20+ model families recognised from filename; correct prompt template applied.
• 📦 Downloaders - Pick popular presets or custom IDs for GGUFs, full HF Transformers snapshots, Ollama models, and llama.cpp binaries without leaving the app.
• 🖧 Local server app - python -m nativelab.server turns any GGUF model into an OpenAI/Anthropic compatible API server. Hardware-aware auto-configuration.

See changelog.txt for the latest release notes and docs/architecture.md for the layered design.

PhonoLab - Android Client

📚 Documentation

The docs are split into short, focused files so you can jump straight to what you need.

Beginner-friendly walkthroughs:

• 🆕 Never used a terminal LLM tool? Start with nativelab/cli/cli_guide.md.
• 🆕 Want to add a lab feature? Read docs/labs.md.

⚡ Quick start

GUI

pip install nativelab
nativelab

The first launch opens the desktop app. Use the Download tab to install llama.cpp binaries, grab a GGUF model, pull an Ollama model from an already-running Ollama daemon, or download a full HF Transformers snapshot. The HF Transformers downloader includes an in-app library installer for the optional Transformers runtime packages. For gated/private Hugging Face repos, sign in from Settings > Accounts > Hugging Face > Login with Hugging Face first, then accept or request access on the repo page if Hugging Face still returns 403.

CLI

pip install nativelab
nativelab --cli

The CLI runs an interactive wizard the first time:

1. Verifies llama-server / llama-cli are present (or guides you to install them).
2. Lets you pick or download a GGUF model from HuggingFace.
3. Asks for a context size with sensible defaults.
4. Opens the terminal control center with Chat, Models, API Models, Skills, Labs, Pipelines, Integrations, Status, and Setup.

nativelab --cli models list
nativelab --cli api-models list
nativelab --cli skills chat-on
nativelab --cli endpoint /snapshot --json
nativelab --cli chat

Full beginner walkthrough: nativelab/cli/cli_guide.md.

🧪 Labs - the experimentation layer

The nativelab/labs/ package is a sandbox for new features. Every lab panel receives a single LabEndpoints instance and uses it for all engine interaction:

from nativelab.labs import LabEndpoints

# Read state
endpoints.status_text     # "🟢 Server  :8612"
endpoints.model_path      # "/abs/path/to/mistral-7b.Q4_K_M.gguf"
endpoints.snapshot()      # {model_name, ctx_value, server_port, …}

# Synchronous LLM call - auto-routes API > server > CLI
endpoints.call_llm(messages=[...], system_prompt="…")

# Reverse routing - ask the host app to change state
endpoints.request_load_model("/path/to/other.gguf")
endpoints.request_context(8192)
endpoints.request_unload()

Add a lab feature by dropping nativelab/labs/<feature>.py with a QWidget panel that has LAB_NAME, LAB_ICON, and a set_endpoints(...) method, then appending it to LAB_FEATURES. Full guide in docs/labs.md.

🛠️ Requirements

• Python 3.10+
• PyQt6 (installed automatically as a dependency)
• llama.cpp binaries - llama-server / llama-cli. The GUI's Download tab installs these for you, or you can drop them in ./llama/bin/.
• Optional: psutil (RAM monitor), pypdf (PDF summarization), pyflakes / flake8 / pylint (CLI lint).
• Optional HF backend: use the HF Transformers downloader's Install Libraries action, or install the displayed Transformers/Torch/safetensors/Accelerate/SentencePiece/Pillow command manually.

Detailed instructions in docs/installation.md.

🤝 Contributing

Issues and PRs welcome. See CONTRIBUTING.md and CODE_OF_CONDUCT.md.

For security disclosures, see SECURITY.md.

📜 License

AGPL v3 - free and open source forever. See LICENSE.

Both NativeLab and PhonoLab are licensed under AGPL v3. NativeLab depends on llama.cpp (MIT) and PyQt6 (GPL/commercial). PhonoLab depends on llama.cpp (MIT) and AndroidX (Apache 2.0).