v4.6

Oobabooga's text-generation-webui, the popular open-source interface for running local large language models like Llama 3 and Mistral, has released version 4.6 with significant upgrades for AI agent functionality and safety. The headline feature is full support for Model Context Protocol (MCP) servers, allowing the interface to connect to external data sources and tools using the same standardized protocol adopted by Anthropic's Claude Desktop and Cursor IDE. This enables local LLMs to access databases, APIs, and filesystems through a unified interface. Additionally, v4.6 introduces granular safety controls with inline approve/reject buttons that appear before each AI tool call executes, giving users precise oversight over when their local models can take actions.

The update also brings substantial UI improvements including independently toggleable sidebars that persist across sessions, better handling of thinking blocks in reasoning models, and security fixes for SSRF vulnerabilities. Under the hood, v4.6 updates core dependencies including llama.cpp to improve performance and adds support for ik_llama.cpp's new quantization types. The portable builds now offer separate packages for CUDA 12.4 and 13.1 environments, with the CUDA 13.1 version being 10% smaller at 686MB for Windows. These changes make text-generation-webui more competitive with commercial AI agent platforms while maintaining its open-source, locally-run philosophy.