# CompresLM

CompresLM is a drop-in LLM proxy that cuts the cost of every LLM call by sending models only the context they need.

## What it is
- OpenAI-compatible and Claude-compatible HTTP proxy.
- Sits between your app and any upstream LLM (OpenAI, Anthropic, or a self-hosted model).
- Compresses bulky context before the request is forwarded — chat history, pasted documents, RAG insertions, and tool/function call results, not just the latest user message.
- Does not store your prompts or provider API keys.
- First 1 million tokens processed are free. After that you pay 20% of dollars saved (or $0.50 per 1M tokens saved for local/unknown-price models).

## Why use it
- Up to 90% lower cost per 1M tokens on compressible workloads.
- On a 200+ case paired public-dataset benchmark, quality is held (no statistically significant drop) on needle-fact, negation, multi-hop and log-search cases, with measured cost reduction of 6.5%–70.4% on those categories.
- Two harder categories (query-focused summarization, and code pasted directly in the user message) are automatically routed through the uncompressed or headroom path instead, because aggressive extraction measurably hurts them — you get the safe result without having to choose.
- Quality often improves because the model is not distracted by irrelevant context.
- Measured to outperform Netflix-style internal head/tail compression tooling on workloads where those tools save ~0% and CompresLM still cuts cost at matched quality.

## Cost per 1M input tokens (measured on this benchmark)
| Model | List price | With CompresLM |
|---|---|---|
| GPT-4o mini | $0.15 | ~$0.084 |
| GPT-4o | $2.50 | ~$1.41 |
| Claude Sonnet | $3.00 | ~$1.69 |
| Self-hosted / local | n/a | billed at $0.50 per 1M tokens saved |

## How to connect (one line)
Point your existing OpenAI/Anthropic SDK at CompresLM:

- base_url = https://compreslm.com/v1
- api_key = your CompresLM token (from the dashboard)
- header X-LLM-API-Key = your upstream provider key (never stored by CompresLM)
- optional header X-LLM-API-Base = custom upstream base URL for self-hosted LLMs

Example (Python OpenAI SDK):

```
from openai import OpenAI
client = OpenAI(
    base_url="https://compreslm.com/v1",
    api_key="YOUR_COMPRESLM_TOKEN",
    default_headers={"X-LLM-API-Key": "YOUR_PROVIDER_KEY"},
)
client.chat.completions.create(model="gpt-4o-mini-smart", messages=[...])
```

Recommended model alias: gpt-4o-mini-smart (auto-routes compression strategy per request).

## Where compression applies
- Chat history and pasted documents in user/system messages.
- RAG-inserted context blocks.
- Tool / function call results (role="tool"), where LLM agent workloads usually carry the most removable bulk.
- Works the same way inside multi-step agent frameworks (e.g. LangGraph) as long as each LLM call in the chain goes through this proxy — every hop gets compressed independently, so long tool-call chains save at every step, not just the first one.

## Pricing
- Free tier: 1,000,000 tokens processed per account.
- After free tier, known-price models: commission = 20% × (cost_without_compression − cost_with_compression).
- Local / unknown-price models: $0.50 per 1,000,000 tokens saved.
- You never pay more than you save on known-price models.
- On-prem and custom plans: contact via the form on the site (we help set up for free).

## Privacy
CompresLM processes each request and forwards it. Provider keys arrive per-request via header and are not persisted. Prompts are not kept as a product dataset. Full detail: /app/privacy.

## Operator
CompresLM is operated by Individual Entrepreneur Mikita Valkunovich, registered in Georgia (identification number 300412284). Full terms: /app/terms.

## Links
- Landing: https://compreslm.com/
- App / signup: https://compreslm.com/app
- Docs: https://compreslm.com/app/docs
- Billing explained: https://compreslm.com/app/billing
- Terms of Service: https://compreslm.com/app/terms
- Privacy Policy: https://compreslm.com/app/privacy
- This file: https://compreslm.com/llms.txt