@ben_labs/semanticzip
v1.0.0
Published
Compress prompts with semantic token compression (SemanticZip) before sending to the LLM — save tokens and reduce cost.
Maintainers
Readme
SemanticZip Token Compression — OpenClaw Plugin
Automatically compress prompts before they reach the LLM using SemanticZip (Semantic Token Compression). Save tokens, reduce costs, and speed up responses — with zero changes to your workflow.
What it does
- Intercepts prompts via the
before_agent_starthook - Sends system prompts and user messages to the SemanticZip compression API
- Removes low-entropy tokens that don't contribute to meaning
- The LLM receives a shorter prompt and responds normally
- Typical savings: 30–70% of input tokens
Install
openclaw plugins install @semanticzip/semanticzipOr install from a local directory:
openclaw plugins install ./path/to/semanticzipConfiguration
Add to ~/.openclaw/openclaw.json:
{
"plugins": {
"entries": {
"semanticzip": {
"enabled": true,
"config": {
"apiKey": "YOUR_API_KEY_HERE",
"threshold": -2.0
}
}
}
}
}Or set the SEMANTICZIP_API_KEY environment variable.
Config options
| Option | Type | Default | Description |
|--------|------|---------|-------------|
| apiUrl | string | https://api.semanticzip.com | SemanticZip API base URL |
| apiKey | string | — | Your SemanticZip API key |
| threshold | number | -2.0 | Compression aggressiveness: -0.5 light, -1.0 moderate, -2.0 aggressive, -3.0 maximum |
| enabled | boolean | true | Enable/disable compression |
| compressSystemPrompt | boolean | true | Compress the system prompt |
| compressUserMessages | boolean | true | Compress user messages |
| minTokensToCompress | number | 50 | Skip messages shorter than this |
Get an API key
Sign up at semanticzip.com to get your API key.
How SemanticZip works
SemanticZip uses a small language model to score every token by its information content (entropy). Tokens that are predictable from context — articles, filler words, redundant punctuation — are removed. The result reads like shorthand but retains full meaning for the LLM.
Example:
- Before:
The quick brown fox jumps over the lazy dog in the park - After:
quick brown fox jumps over lazy dog park - Savings: ~40%
License
MIT
