What is omniglyph?

OmniGlyph is a tool that helps reduce the cost of using AI models like Claude when working with large amounts of text context. The core idea is simple: instead of sending big chunks of text, such as system prompts, tool documentation, or old conversation history, to the AI as plain text, OmniGlyph turns that text into dense image files, specifically PNG pages. Since AI providers often charge for images based on their size rather than the amount of text inside them, this trick can cut your bill significantly. The project's own measurements show a 59 to 70 percent reduction in end to end cost. OmniGlyph works as a local proxy that sits between your computer and the AI provider. It rewrites bulky parts of your requests into these compressed image pages before anything leaves your machine, so nothing extra is sent elsewhere. The project aims for exact billing math rather than rough guesses, meaning it calculates the real formula each provider uses for image tokens and only converts to an image when doing so actually saves money. Not every AI model can read these dense images accurately. The project includes a fail closed gate, so if a model cannot reliably read the compressed renders, it blocks the conversion rather than risking silent errors or made up answers. According to the README, Claude Fable 5 reads these images with 100 percent accuracy in testing, while some other models such as GPT-5.5 and Gemini 2.5-flash struggle or fail entirely and were blocked by the gate. The project is written mostly in TypeScript and is available as a CLI, a proxy, and a Cloudflare Workers deployment. It is part of a larger project family called OmniRoute, and its documentation has been translated into 42 languages. To get started, users can run it directly with npx and point their Claude Code setup at the local proxy address. This project would mainly interest developers and technical users who already work with AI coding assistants and want to reduce their token costs, particularly people running long coding sessions with large context windows.