🧠 The Problem

Every token you send to an AI model costs money. Long, verbose prompts with filler words, redundancy, and conversational fluff can double your API bill without adding any value to the response.

Most developers don't realize that up to 40% of their prompt tokens are wasted on words the model doesn't need to understand your intent.

Ayu-Gravity solves this. It sits between you and the AI — compressing your prompts on-the-fly using a fast, cheap model before forwarding them to your primary expensive model. You type naturally. The AI receives a lean, optimized prompt. You save tokens. You save money.

✨ Features

🔌 Supported Providers

🎯 Supported Models

Anthropic

OpenAI

Google Gemini

Ollama

Models are fetched dynamically from your local Ollama installation. Whatever you have pulled locally — Llama, Mistral, Phi, Qwen — it appears automatically in the dropdown.

GitHub Copilot

Model selection is automatic via the VS Code Language Model API. No configuration needed.

🔧 How It Works

┌──────────────┐     ┌──────────────────┐     ┌──────────────┐
│  You type a  │────▶│  Ayu-Gravity     │────▶│  AI Model    │
│  prompt      │     │  compresses it   │     │  responds    │
│  (100 tokens)│     │  (65 tokens)     │     │  (streamed)  │
└──────────────┘     └──────────────────┘     └──────────────┘

1. You type your prompt naturally in the sidebar chat
2. Token estimation — if your prompt is ≥ 20 estimated tokens, compression kicks in
3. Compression — a fast, cheap model (Haiku / gpt-4o-mini / Flash) strips filler words, redundancy, and verbose phrasing in under 10 seconds
4. Original + compressed — you see both in the UI with a toggle
5. Forwarding — the compressed prompt gets sent to your selected model
6. Streaming — the response streams back token-by-token with live Markdown rendering
7. Savings — you save 30–50% of input tokens on most prompts

📦 Installation

From VS Code Marketplace

1. Open VS Code
2. Go to Extensions (Ctrl+Shift+X)
3. Search for "Ayu-Gravity"
4. Click Install

From VSIX File

code --install-extension ayu-gravity-2.1.0.vsix

🔑 Setting Up API Keys

1. Open the Command Palette (Ctrl+Shift+P)
2. Run "Antigravity: Set API Key"
3. Select your provider (Anthropic / OpenAI / Gemini)
4. Paste your API key (input is hidden)
5. Done! The provider now appears in your dropdown

Note: API keys are stored in VS Code's native SecretStorage — encrypted at rest and never written to disk in plaintext.

No API Key Needed For:

• GitHub Copilot — works automatically if you have a Copilot subscription
• Ollama — runs locally, just have the Ollama server running on localhost:11434

🗜️ Compression Modes

Tip: Compression is only applied to prompts with 20+ estimated tokens. Short prompts are sent as-is.

📸 Screenshots

Coming soon — screenshots of the sidebar panel, model selector, and compression in action.

❓ FAQ

Q: Does compression affect response quality? A: No. Compression removes filler words and redundancy — not meaning. LLMs actually perform better with concise, clear prompts.

Q: What model is used for compression? A: The cheapest, fastest model for each provider: claude-3-5-haiku for Anthropic, gpt-4o-mini for OpenAI, gemini-1.5-flash for Gemini. Compression cost is negligible.

Q: Can I see what was compressed? A: Yes! Every compressed message has a "View compressed prompt" toggle in the UI that shows you exactly what was sent.

Q: Does it work offline? A: Only with Ollama (local models). All other providers require internet access.

Q: Are my API keys safe? A: Yes. All keys are stored in VS Code's native encrypted SecretStorage. They are never logged, transmitted, or written to plaintext files.

Q: What happens if compression fails? A: The original prompt is sent as-is. Compression has a 10-second timeout and graceful fallback — it never blocks your message.

Q: Can I choose which model to chat with? A: Yes! The model selector dropdown lets you pick any available model for your selected provider.

📄 License

MIT License — free for personal and commercial use.