caveman-shrink

MCP middleware. Wrap any MCP server. Cut the prose. Keep the substance.

caveman-shrink is a stdio proxy for the Model Context Protocol. It sits between Claude (or any MCP client) and an upstream MCP server, and compresses the prose fields (description, etc.) using the same boundaries as the caveman skill — preserving code, URLs, paths, and identifiers while stripping articles, filler, hedging, and pleasantries.

The result: tool catalogs that the model burns fewer tokens to read, with no change to tool semantics.

Install

npm install -g caveman-shrink
# or run directly via npx
npx caveman-shrink <upstream-command> [...args]

Use it

Wrap any MCP server in your Claude Code (or other client) config:

{
  "mcpServers": {
    "fs-shrunk": {
      "command": "npx",
      "args": [
        "caveman-shrink",
        "npx", "@modelcontextprotocol/server-filesystem", "/path/to/dir"
      ]
    }
  }
}

The proxy spawns the upstream as a subprocess, intercepts tools/list, prompts/list, resources/list responses, and rewrites the description fields (and anything else you list in CAVEMAN_SHRINK_FIELDS).

What it does NOT touch

By design, v1 is conservative:

• Request bodies going to the upstream are passed through unchanged.
• Tool call responses (tools/call) are passed through unchanged. We don't want to risk silently mutating the data the upstream returns to the model.
• Identifiers, URLs, paths, and code-looking tokens inside any prose are preserved exactly. Same boundaries as the parent caveman skill.

Configuration

| Env var | Default | What | |---|---|---| | CAVEMAN_SHRINK_FIELDS | description | Comma-separated list of field names to compress | | CAVEMAN_SHRINK_DEBUG | 0 | Set to 1 to log per-field compression deltas to stderr |

Status

Pre-1.0 — the compression rules and field set may change. The plugin is part of the caveman ecosystem; see the parent repo for the full skill suite (caveman, cavemem, cavekit, cavecrew, caveman-stats, caveman-init).

License

MIT.