14ms cuts 30K tokens of waste — code-to-LLM toolchain still stuck in cottage industry

What this is

repOx is a Rust-written command-line tool with one job: "pack" a code repository into a prompt you can feed directly to a large language model. Developer WVDYC open-sourced the project (GitHub: WVDYC/repOx) in the Reddit LocalLLaMA community. Its two selling points are speed and cleanliness. It processes a 3,000-file repository in just 14 milliseconds, and by default strips out "prompt junk" like Cargo.lock, package-lock.json, SVGs, and binary files — in ordinary packers, these routinely burn through 30,000+ tokens (the smallest unit text is broken into; models bill by it and use it to size their context window), contributing zero value to a model's understanding of the code. It also bundles offline tokenizers for Claude, GPT, Gemini, DeepSeek, and Llama, so you can tell whether you'll exceed context capacity before sending the prompt.

Industry view

The supportive angle: the open-source community pounced on the discussion because this hits a long-standing pain point for people running models locally. They want to stuff an entire project into DeepSeek R1 or Qwen for code analysis, and watching lockfiles burn tens of thousands of tokens is pure waste. Some developers reported the tool running 200x faster than existing packers.

The skeptical angle: the cooler take is that this functionality is already covered by tools like repomix and aider. WVDYC himself admits it's "customized for my own workflow." What actually deserves our attention is this: vibe coding — directing AI to write code in natural language — has been hot for over a year, yet the toolchain layer remains fragmented. The fact that a 14ms micro-optimization can spark community discussion tells us that Cursor, Copilot, and the other headline players haven't nailed the basics of "code packing." The big names are racing on models; the engineering capability that actually moves output has been left to open-source independents.

Impact on regular people

For enterprise IT: this means that internal tools using LLMs for code review or codebase Q&A will mostly still need to be built in-house or stitched together — don't assume that buying an API means plug-and-play. "Feeding the data" is itself an engineering threshold.

For working professionals: programmers looking to boost productivity with local LLMs will find that "how to feed the code" matters more than "which model to pick." It's a grossly underestimated engineering capability.

For the consumer market: this kind of tool is still far from end users, but it points to a trend — eventually every non-programmer may need something similar to package their own documents, spreadsheets, and knowledge bases for AI.