Skip to content
diegosouzapw edited this page Jul 8, 2026 · 1 revision

OmniGlyph Wiki

Context as Image. OmniGlyph is a local compression proxy for LLMs: it renders the bulky parts of each request — system prompt, tool docs, old history, large tool outputs — as dense PNG pages and lets the model read them. Models bill text per token but bill an image by its dimensions, not by how much text is packed inside. Measured end-to-end bill reduction: 59–70%.

The proxy only converts when the exact per-provider billing math says it wins, and only for models that passed the reading benchmark (fail-closed gate).

Start here

  • Quick Start — install and point Claude Code at the proxy in two commands.
  • How It Works — the transform pipeline and the two gates.
  • Billing Math — the exact per-provider token accounting.
  • Benchmarks — how the savings and read-accuracy are proven.
  • OmniRoute Integration — running as the native engine.
  • FAQ — the honest part: limits, what it never does.

At a glance

Package omniglyph on npm
Runtime Node ≥18 · Cloudflare Workers
License MIT
Providers with exact math Anthropic · OpenAI · Gemini
Approved for imaging today Claude Fable 5 (100% exact reads)
Docs English + 41 translated languages

Part of the OmniRoute family — the free AI gateway.

Clone this wiki locally