gwen-digestor

Model Context Protocol server for conversation compression.

Reduces token consumption by compressing conversation exchanges before they enter the LLM context window. Uses deterministic, embedding-free compression — no external APIs, no GPU required.

Features

• 4 MCP tools: digest_input, compress_response, cache_reference, session_stats

• Mode-aware compression: auto-detects checkin, task, narrative, or casual conversation

• Content-type detection: smart JSON crushing, code comment stripping, prose pass-through

• Gzip-compressed reference cache: SQLite-backed key-value store with TTL expiry

• Token savings tracking: persistent stats across sessions

📊 View the Token Reduction Report — a professional breakdown with compression metrics and visual charts.

Related MCP server: Memory Cache Server

Compression Levels

Tools

digest_input

Compresses incoming messages by mode. Strips conversational filler, extracts health metrics in checkin mode, removes boilerplate in task mode.

compress_response

Compresses outgoing responses with mode-aware sentence truncation.

cache_reference

Gzip-compressed key-value store for reference texts. Configurable TTL (default 24h).

session_stats

Real-time token savings dashboard showing compression rates across all calls.

Installation

pip install mcp fastmcp

Usage

Register as an MCP server in your client config:

{
  "mcpServers": {
    "gwen-digestor": {
      "command": "python3",
      "args": ["/path/to/gwen_digestor.py"],
      "transport": "stdio"
    }
  }
}

Then call the tools from your LLM session:

digest_input("hey, just checking in — slept okay, pain 3/10 today, stress 5/10")
→ [MODE:checkin@25%] SLEEP:okay|PAIN:3/10|STRESS:5/10

Storage

• Cache DB: ~/.gwen-digestor/cache.db (SQLite, gzip-compressed blobs)

• Stats: ~/.gwen-digestor/stats.json (persistent across sessions)

• Dependencies: Python 3.10+, mcp, fastmcp

License

MIT