AIP Proxy

Token compression proxy for LLM APIs. Reduce your AI coding costs by 15-40% without losing quality.

AIP Proxy sits between your AI IDE (Antigravity, Cursor, VS Code, etc.) and the LLM API, compressing prompts on the fly before they reach the model.

How it works

Your IDE  ──>  AIP Proxy (localhost:8090)  ──>  OpenAI / Gemini / Claude API
                    │
                    ├─ Whitespace normalization
                    ├─ Code comment removal
                    ├─ Block deduplication
                    └─ Pattern abbreviation

4 compression passes, configurable by level:

Level	Passes	Typical savings
0	None (passthrough)	0%
1	Whitespace normalization	5-10%
2	+ Code compression + deduplication	15-25%
3	+ Pattern abbreviation	25-40%

Install

pip install aip-proxy

Quick start

# Start proxy targeting OpenAI
aip-proxy start --target https://api.openai.com/v1 --port 8090

# Or targeting Google Gemini
aip-proxy start --target https://generativelanguage.googleapis.com --port 8090

# Or any OpenAI-compatible API
aip-proxy start --target https://api.anthropic.com --port 8090

Then change your IDE's API endpoint to http://localhost:8090/v1.

Usage with Antigravity

1. Install: pip install aip-proxy
2. Start: aip-proxy start --target https://generativelanguage.googleapis.com --port 8090
3. In Antigravity settings, set API endpoint to http://localhost:8090
4. Done — you'll see savings in the proxy stats

Usage with Cursor / VS Code

1. Install: pip install aip-proxy
2. Start: aip-proxy start --target https://api.openai.com/v1 --port 8090
3. In your IDE settings, change the API base URL to http://localhost:8090/v1
4. Keep your API key as usual — the proxy forwards it transparently

Options

aip-proxy start --help

Options:
  --target, -t    Target API URL (required)
  --port, -p      Port to listen on (default: 8090)
  --host          Host to bind (default: 127.0.0.1)
  --level, -l     Compression: 0=off, 1=light, 2=balanced, 3=aggressive (default: 2)
  --no-cache      Disable response caching
  --cache-ttl     Cache TTL in seconds (default: 300)

Endpoints

Endpoint	Description
GET /health	Proxy status and basic stats
GET /stats	Detailed compression and cache statistics
* /{path}	Proxied to target API

Python API

from aip_proxy import TokenCompressor

tc = TokenCompressor(level=2)
messages = [
    {"role": "user", "content": "your long prompt here..."}
]
compressed = tc.compress_messages(messages)
print(tc.get_savings())
# {'original_chars': 1500, 'compressed_chars': 1100, 'saved_chars': 400, 'savings_pct': 26.7, 'calls': 1}

How does it save money?

LLM APIs charge per token. A typical coding session sends thousands of tokens in context — much of it is:

• Redundant whitespace and blank lines
• Comments in code blocks (the model doesn't need them)
• Repeated code blocks across messages
• Verbose filler phrases

AIP Proxy removes this noise while preserving the semantic content the model needs.

License

Copyright (c) 2024-2026 Carmen Esteban. All rights reserved. No part of this software may be copied, modified, distributed or used without express written permission.