v4.2
The popular local LLM interface now lets you run Claude Code and Cursor locally with full API compatibility.
Oobabooga has launched version 4.2 of text-generation-webui, the massively popular open-source interface for running large language models locally. The headline feature is a new Anthropic-compatible API endpoint (/v1/messages), which allows developers to connect clients like Claude Code and Cursor directly to locally-hosted models. This endpoint supports system messages, content blocks, tool use, image inputs, and thinking blocks, effectively letting users run Anthropic-style applications on their own hardware. The update also includes a complete visual overhaul with new colors, borders, and button styles across both light and dark modes.
Under the hood, significant improvements target model training and efficiency. Gradient checkpointing is now enabled by default, reducing VRAM usage during training. The team removed 52 obsolete instruction templates from 2023 (including Airoboros and Vicuna v0) and reorganized the training UI for better usability. For API users, the default sampling parameter `top_p` has been set to 0.95. The update also bundles numerous bug fixes addressing issues from prompt corruption in chats to dependency resolution for AMD ROCm installations, making the platform more stable for a wide range of hardware setups.
- Adds full Anthropic-compatible API endpoint (/v1/messages) for running Claude Code and Cursor locally
- Completely redesigned UI theme with updated colors and styles for light/dark modes
- Enables gradient_checkpointing by default for lower VRAM usage during model training
Why It Matters
This democratizes access to advanced AI workflows, letting developers run and test Claude-like applications entirely offline on their own machines.