LatentGate — VSCode Extension

Local-first AI compression for VSCode. Compress images, text, conversations, and documents locally via Ollama before sending to cloud LLMs. Save ~80% on token costs.

Features

• Compress Images — Right-click any image → LatentGate compresses it to ~150 tokens
• Compress Text — Select text → Ctrl+Shift+Alt+C to compress
• Cost Dashboard — Activity bar panel showing token savings and stats
• MCP Integration — Auto-configures MCP server for Copilot Chat
• Status Bar — Live token savings counter

Prerequisites

1. Python 3.10+ with latent-gate installed:

   pip install latent-gate[all]
   

2. Ollama running with required models:

   ollama pull llava:7b
   ollama pull llama3:8b
   

Commands

Configuration

Open VSCode Settings → search latentGate to configure:

• latentGate.visionModel — Vision model (default: llava:7b)
• latentGate.predictorModel — Text model (default: llama3:8b)
• latentGate.remoteProvider — Remote LLM provider
• latentGate.remoteModel — Remote model name
• latentGate.selectiveDecoding — Skip API calls for unchanged scenes

MCP Server

The extension auto-configures the MCP server for VSCode Copilot Chat. This enables Copilot to use LatentGate's compression tools automatically.

Manual Configuration

Add to your settings.json:

{
  "copilot.chat.mcpServers": {
    "latent-gate": {
      "command": "python",
      "args": ["-m", "latent_gate.mcp_server"]
    }
  }
}

How It Works

1. Local Processing — Images/text are processed locally via Ollama (free)
2. Compression — Extracts essential semantics (~150 tokens vs ~1200 raw)
3. Remote Decoding — Sends compact payload to cloud LLM
4. Savings — ~80% reduction in token costs

Requirements

• Python 3.10+
• Ollama (for local processing)
• VSCode 1.85+

Links

• GitHub
• PyPI
• Documentation

License

MIT License