mandiproxy

Route LLM API calls through Mandarin Chinese to reduce token usage by 30–60%.

How it works

Chinese characters encode more information per token than English words. A BPE tokenizer (used by GPT, Claude, etc.) assigns roughly 1–3 tokens per Chinese character, but each character can represent a full word or concept. The same semantic content in Mandarin uses significantly fewer tokens.

mandiproxy is a local HTTP proxy that:

1. Intercepts your LLM API call
2. Translates message content to Mandarin before sending
3. Receives the Mandarin response
4. Translates back to English
5. Returns it to your tool — transparently

Quick start

pip install mandiproxy

# Start the proxy
mandiproxy start

# In another terminal, point your tool at the proxy:
ANTHROPIC_BASE_URL=http://localhost:8765 claude        # Claude Code
OPENAI_BASE_URL=http://localhost:8765/v1 codex         # OpenAI Codex CLI

CLI

# Test a single prompt and see token savings
mandiproxy chat "Explain how neural networks work" --show-zh

# View session metrics
mandiproxy metrics

Environment variables

Variable	Default	Description
MANDIPROXY_UPSTREAM_OPENAI	https://api.openai.com/v1	Upstream OpenAI-compatible URL
MANDIPROXY_UPSTREAM_ANTHROPIC	https://api.anthropic.com/v1	Upstream Anthropic URL
MANDIPROXY_TRANSLATION_MODEL	gpt-4o-mini / claude-haiku-*	Model used for translation

Agentic mode

Works with any tool that supports a custom base URL:

# Claude Code
export ANTHROPIC_BASE_URL=http://localhost:8765
claude

# OpenAI Codex
export OPENAI_BASE_URL=http://localhost:8765/v1
codex

Metrics

Every response includes a mandiproxy_metrics field:

{
  "mandiproxy_metrics": {
    "input_savings_pct": 42.3,
    "output_savings_pct": 38.1,
    "total_savings_pct": 40.5,
    "session_total_savings_pct": 39.2
  }
}