TextGen becomes portable desktop app, challenges LM Studio with privacy and tools
Open-source LLM launcher now ships no-install builds for all platforms, zero telemetry.
TextGen (formerly text-generation-webui, also known as oobabooga) has evolved from a web UI to a native desktop app after two years of development. The new Electron-based portable build eliminates installation: download, unzip, double-click. All chat histories and settings live in a bundled `user_data` folder. Crucially, TextGen is privacy-first—unlike LM Studio, it phones home with zero telemetry. Builds are available for CUDA, Vulkan, CPU-only, Mac (Apple Silicon and Intel), and ROCm.
The app ships ik_llama.cpp, a custom fork offering state-of-the-art quant types (IQ4_KS, IQ5_KS). It supports tool-calling through three mechanisms: simple .py tools, HTTP MCP servers, and stdio MCP servers, each with optional approve/reject confirmations. Built-in web search via ddgs works as a tool or a checkbox, and PDF extraction uses PyMuPDF. The API is strictly OpenAI/Anthropic-compatible—works with Claude Code by setting `ANTHROPIC_BASE_URL`. Custom characters for casual chats, Jinja2 template rendering (no crashes), and trafilatura for clean web page fetching round out the feature set. The project remains free, open-source (AGPLv3), and a passion hobby for its creator.
- Portable builds for all platforms: no install, no traces outside the extracted folder, zero outbound requests.
- Includes ik_llama.cpp with advanced quant types (IQ4_KS, IQ5_KS) for superior accuracy.
- Tool-calling with confirmations, support for MCP servers, and full OpenAI/Anthropic API compliance (works with Claude Code).
Why It Matters
Gives professionals a fully private, customizable, offline LLM platform with advanced tooling, rivaling paid services.