llm-telephone

Pass text through a chain of 20 LLM translations and watch the meaning drift.

What is this?

llm-telephone is a command-line tool that feeds a piece of text through a configurable chain of language-to-language translations via the OpenRouter API, then brings it back to your chosen return language.

Like the childhood game of "telephone", each translation step introduces small distortions. After 20 hops through English → Japanese → French → German → … → Simplified Chinese, you can observe how meaning drifts, idioms break down, and cultural context gets lost.

Why is it interesting?

The original poetic exclamation (龙，可是帝王之征啊) loses its rhetorical flourish and emerges as a dry factual statement — same core meaning, completely different register.

Installation

pip install llm-telephone

Requirements

• Python 3.10 or newer
• An OpenRouter API key

Configuration

Set your OpenRouter API key in the environment before running:

export OPENROUTER_API_KEY="sk-or-..."

Alternatively, OPENROUTER_KEY is also accepted.

Quick start

# Translate a single phrase through the default 20-language chain
llm-telephone "The quick brown fox jumps over the lazy dog"

# Read from a file instead
llm-telephone --text-file origin.txt

# Only run 5 rounds
llm-telephone "Hello world" --rounds 5

# End the chain in English instead of Simplified Chinese
llm-telephone "Hello world" --rounds 10 --return-lang English

# Output structured JSON (useful for piping / analysis)
llm-telephone "Hello world" --rounds 5 --output-format json

# Save logs to files
llm-telephone "Hello world" \
    --process-output process.txt \
    --final-output final.txt

All options

Usage: llm-telephone [OPTIONS] [TEXT]

  Pass TEXT through a chain of LLM translations and watch the meaning drift.

Options:
  --model TEXT                    OpenRouter model ID.  [default: google/gemini-3-flash-preview]
  --rounds INTEGER                Number of translation rounds.  [default: 20]
  --text-file PATH                Read initial text from a UTF-8 file.
  --process-output TEXT           Write per-round log to this file path.
  --final-output TEXT             Write the final text to this file path.
  --output-format [text|json]     Output format.  [default: text]
  --return-lang TEXT              Language to end the chain with.  [default: Simplified Chinese]
  --version                       Show the version and exit.
  -h, --help                      Show this message and exit.

JSON output format

When --output-format json is used, the tool writes a single JSON object to stdout:

{
  "model": "google/gemini-3-flash-preview",
  "rounds": 5,
  "return_lang": "Simplified Chinese",
  "initial_text": "Hello world",
  "final_text": "你好，世界",
  "chain": [
    {
      "round": 1,
      "target_language": "English",
      "input": "Hello world",
      "output": "Hello, world!",
      "elapsed_s": 0.823
    }
  ]
}

Python API

You can also use llm-telephone as a library:

from llm_telephone import run_chain
from llm_telephone.api_client import OpenRouterClient
from llm_telephone.languages import get_chain

client = OpenRouterClient(api_key="sk-or-...", model="google/gemini-3-flash-preview")
records = run_chain("Hello world", client, rounds=5, output_format="json")
for r in records:
    print(f"Round {r['round']:02d} [{r['target_language']}]: {r['output']}")

Development

See CONTRIBUTING.md for setup instructions.

License

MIT