Compressor Reflex MCP

Compressor Reflex MCP is a Model Context Protocol (MCP) server and transparent proxy that provides high-fidelity tool-output compression for Cursor, Antigravity IDE, Claude Desktop, and other MCP-compliant developer environments.

Powered by aialchemist-dev/compressor-reflex (fine-tuned ModernBERT-151M), the model reduces tool-output token consumption by up to 90% while guaranteeing 100% retention of compiler errors, test failures, target locations, and decisive anchors.

Key Capabilities

• 89.7% Measured Tool Compression: Condenses extensive test suites, git diffs, directory listings, and file dumps into their essential operational lines.

• 100.0% Critical Anchor Retention: Calibrated at decision threshold tau* = 0.50 across held-out evaluations. Error tracebacks, fail signatures, and query targets remain intact.

• Fail-Open Bypass Policy: Outputs containing <= 5 physical lines or <= 64 tokens automatically bypass compression and pass through verbatim, avoiding latency overhead on short outputs.

• Dual Deployment Modes:

  1. Direct MCP Tools: Exposes standard callable tools (compress_tool_output, compress_file) for explicit agent invocation.

  2. Transparent Proxy: Wraps any standard MCP server (such as filesystem, git, or terminal) to automatically compress output streams before delivery to the context window.

• Automated Weight Management: Model weights (INT8 ONNX) and tokenizers are automatically retrieved from Hugging Face Hub on initial startup and cached locally.

Related MCP server: genpark-agent-context-dynamic-compressor-skill

Installation

From PyPI

pip install compressor-reflex-mcp

Or run directly without installation via uvx:

uvx compressor-reflex-mcp serve

From Source or Git

git clone https://github.com/ericmaddox/compressor-reflex-mcp.git
cd compressor-reflex-mcp
pip install -e .

Optional Model Pre-Caching

To download the model weights ahead of time:

compressor-reflex-mcp download

Model files are cached in the standard user cache directory (~/.cache/compressor-reflex/ or %LOCALAPPDATA%/compressor-reflex/). The cache path can be overridden using the COMPRESSOR_MODEL_DIR environment variable.

IDE Configuration

Cursor

Add the server definition to your workspace .cursor/mcp.json or global Cursor settings:

{
  "mcpServers": {
    "compressor-reflex": {
      "command": "python",
      "args": ["-m", "compressor_reflex_mcp.server"]
    }
  }
}

Or using uvx:

{
  "mcpServers": {
    "compressor-reflex": {
      "command": "uvx",
      "args": ["compressor-reflex-mcp", "serve"]
    }
  }
}

Antigravity IDE

Add to ~/.gemini/config/mcp_config.json or your project .gemini/mcp_config.json:

{
  "mcpServers": {
    "compressor-reflex": {
      "command": "python",
      "args": ["-m", "compressor_reflex_mcp.server"]
    }
  }
}

Transparent Proxy Mode (Antigravity and Cursor)

Wrap existing tools to automatically compress outputs from heavy providers (for example, filesystem or git inspection):

{
  "mcpServers": {
    "filesystem-compressed": {
      "command": "compressor-reflex-mcp",
      "args": [
        "proxy",
        "--",
        "npx",
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "."
      ]
    }
  }
}

Claude Desktop

Update claude_desktop_config.json:

{
  "mcpServers": {
    "compressor-reflex": {
      "command": "python",
      "args": ["-m", "compressor_reflex_mcp.server"]
    }
  }
}

Tool Reference

Command-Line Interface

Inputs exceeding 5,242,880 characters or 10,000 lines bypass model work and return the original text without truncation. For these resource-limit bypasses, token counts are null because tokenization is skipped; the compression ratio is zero and MCP metadata includes a bypass_reason. For in-budget inputs, selected lines are returned in full even when their scoring tokens are shortened to fit the model sequence budget.

# Launch the stdio MCP server
compressor-reflex-mcp serve

# Run as transparent proxy wrapping another command
compressor-reflex-mcp proxy -- npx -y @modelcontextprotocol/server-filesystem /path/to/project

# Compress a file or standard input directly
compressor-reflex-mcp compress tests/results.log --intent "find failures"

# Verify model cache and runtime configuration
compressor-reflex-mcp info

# Pre-fetch weights from Hugging Face
compressor-reflex-mcp download

Empirical Performance

Metrics collected across real multi-turn developer sessions in IDE environments:

Agent Instructions

For system prompt guidelines and autonomous agent behavior rules, refer to AGENTS.md.

License

This project is licensed under the MIT License. See LICENSE for details.
Model architecture and pre-trained weights are hosted at Hugging Face.