Aphrodite

CCR compression proxy + absorptive preview pipeline for Hermes Agent. Up to 610× compression on the standard corpus (132× overall), ~10 ms end-to-end, type-aware classifier, TOML-driven, opt-in config auto-reload.

One binary. Zero dependencies. Millions of tokens saved.

The problem

Every file read, build, code search, or browser open floods the agent's context with raw output - compilation logs, accessibility trees, JSON blobs. The agent spends its budget reading noise instead of reasoning.

Aphrodite intercepts output before it reaches the LLM and replaces it with a compact, structured preview. The agent sees ~15 tokens of metadata instead of hundreds - and retrieves the full content only when it actually needs it.

The pipeline

1. 1. Classify. A type-aware classifier names the content (build output, diff, JSON, browser snapshot...). Classification alone takes 40-123 ns.
2. 2. Preview. An enriched, type-aware preview is rendered.
3. 3. Store. Full content lands in CCR storage (BLAKE3, SQLite or in-memory) behind a <<<CCR:hash|type|size>>> marker.
4. 4. Decide. The agent retrieves only when needed, via aphrodite_retrieve(hash).

Aphrodite reaches the agent two ways: as a native Hermes plugin with hook-level interception (no API round-trip), or as a reverse proxy for any OpenAI-compatible client, compressing Chat Completions responses through CCR with a tool relay.

The numbers

Measured in the real world across 200 Hermes sessions (6,843 API calls, Sep 18-20, 2026): 156,342,817 tokens saved of 581,302,405 expanded - 26.9% (the exact percentage is a band: ~24-33% across a 3.0-4.5 chars/token sensitivity range, centered ~27%). 183 of 200 sessions saved; tool-heavy workloads scale to 20-25%, pure-tool to 76-82%.

Install

The recommended path is the Hermes plugin; the proxy binary also installs standalone via Cargo. Full instructions on the setup page.

git clone https://github.com/PlayForm/Aphrodite-Hermes.git
cd Aphrodite-Hermes
bash download.sh
hermes plugins enable aphrodite