Un desarrollador ha conseguido reducir un 70% los tokens de Claude Code

Summary

The video introduces PXPipe, an open-source local proxy designed to reduce Claude Code’s token usage—and therefore its cost—without changing the model or summarizing the conversation.

How it works

Claude Code normally sends a large request containing the system prompt, tool descriptions and schemas, conversation history, and tool results. PXPipe sits between Claude Code and the model API, selectively converting bulky content into dense images. Because image processing is priced differently from text tokens, the same information may cost fewer tokens when sent as an image.

PXPipe’s process includes:

• Separating relatively static system-prompt content from dynamic content, then combining static prompt material with tool documentation into a “static slab.”
• Replacing lengthy tool descriptions with short pointers to an image-based tool reference.
• Converting large tool results and older, completed parts of conversation history into images. The most recent four messages remain text.
• Estimating whether image conversion is cheaper than sending content as text, and leaving it as text when conversion is not worthwhile.
• Adding a plain-text factsheet of important exact values—such as IDs, URLs, paths, and email addresses—that could be misread in an image.

PXPipe works best with models capable of reading dense text in images. It also limits how much tool output it converts: when a result exceeds its image limit, PXPipe keeps the beginning and end and inserts a notice about the omitted content.

Reported results and limitations

The presenter cites developer benchmarks in which PXPipe reduced request size by about 65% while solving the same 10 software-engineering exercises as Claude Code using text. In a harder test, PXPipe solved 14 of 19 exercises, compared with Claude Code’s 15 of 19, while reducing request size by about 60%.

A separate demonstration reportedly reduced a 400,000-token session’s cost from $42 to $6.60. In the presenter’s own test, a simple request went from about 134,000 tokens to 70,000—a 48% reduction—and the estimated cost fell from $1.34 to $0.70. The video’s broader savings claim is 59–70%, though the live example falls short of that range.

The main trade-off is reliability: very small text in images may be misread, especially exact values. PXPipe’s factsheet is intended to protect critical data, but image conversion may still introduce errors. The tool is also model-dependent, so the presenter recommends testing it on your own work.

Setup guide shown

The tutorial demonstrates a Windows setup using two terminal windows:

1. Install Node.js 18 or later and Claude Code, then sign in.
2. Install PXPipe and start it in one terminal.
3. Use the local proxy connection command in a second terminal, then launch Claude Code through that connection.
4. Open PXPipe’s local dashboard to inspect requests, images, token reductions, costs, and errors.

With this setup, PXPipe must be started manually. A simpler Warp method is mentioned, but it was not available in the stable version used in the video. The presenter also says the demonstrated model was not included with the Claude Pro plan and required additional funds.

Main speaker and sources

• Main speaker: Alejandro Pardo, from the Alejandro Pardo | IA Automatización channel.
• Other sources referenced: PXPipe’s developers, for the published benchmarks and tool implementation; the video also shows a creator demonstration and the presenter’s own live test.