Developer builds warpdrv, an open-source local LLM coding harness using Qwen 27B
A developer has released warpdrv, a free, AGPL-licensed local LLM harness and coding assistant built primarily using a local Qwen 3.x 27B model. The tool features Just-in-Time (JIT) code reviews, nested sub-agent chats, built-in Model Context Protocol (MCP) tools, and voice dictation support. This project demonstrates the viability of using local, medium-sized open-source LLMs like Qwen 27B to build complex software development tools without relying on cloud-based APIs. It provides developers with a privacy-focused, highly customizable alternative to commercial coding assistants like Claude Code or OpenCode. Warpdrv supports Windows and Linux, runs GPU-agnostically (CUDA, ROCm, Vulkan) via custom-compiled llama.cpp backends, and includes an adversarial AI reviewer to check messages. The author developed the tool on a high-end local setup featuring a Strix Halo 128GB system and an RTX Pro 5000 48GB GPU.
## BACKGROUND
Model Context Protocol (MCP) is an open-source standard introduced by Anthropic that allows AI models to connect securely to external data sources and tools. Meanwhile, llama.cpp is a highly popular, high-performance C/C++ inference engine designed to run LLMs locally in the GGUF format. OpenCode is an existing open-source AI coding agent that runs locally in the terminal or IDE.