- name
- 3-Layer Token Compressor — Cut AI API Costs 40-60%
- description
- Pre-process prompts through 3 compression layers before sending to paid APIs. Uses a local Ollama model to intelligently compress messages and summarize history. Same quality, fewer tokens, lower bills.
- author
- @TheShadowRose
- version
- 1.1.0
- tags
- ["token-optimization", "cost-reduction", "compression", "api-costs", "prompt-optimization", "budget"]
- license
- MIT
3-Layer Token Compressor — Cut AI API Costs 40-60%
Pre-process prompts through 3 compression layers before sending to paid APIs. Uses a free local Ollama model to do the compression work — your paid API only sees the condensed result.
Runtime Requirements
| Requirement | Details |
|---|---|
| Ollama | Must be running locally (default: localhost:11434) |
| Local model | A small model for compression (e.g. llama3.1:8b). Configurable via compressionModel option. |
| Node.js | 14+ |
Ollama is required at runtime. The compressor sends prompts to your local model — not to any external API.
What This Skill Sends to the Local Model
This skill sends the following to your local Ollama model:
| Operation | System prompt | User prompt |
|---|---|---|
| Message compression | You are a text compression tool. Output only what is asked, nothing else. | Your message + instruction to compress |
| History summarization | Same | Old conversation turns + instruction to summarize |
No data is sent to external APIs. All compression happens locally.
Side Effects
| Type | Description |
|---|---|
| NETWORK | HTTP to localhost:11434 only — your local Ollama instance |
| MEMORY | Response cache stored in-memory (Map, configurable size/TTL) |
| DISK | None — cache is not persisted to disk |
Setup
const TokenCompressor = require('./src/token-compressor');
const compressor = new TokenCompressor({
ollamaHost: 'localhost', // default
ollamaPort: 11434, // default
compressionModel: 'llama3.1:8b', // default — any Ollama model works
maxUncompressedTurns: 10, // keep last N turns verbatim
cacheMaxSize: 100,
cacheTTL: 3600000 // 1 hour
});See README.md for full API documentation and usage examples.