Project: promptdry — an MCP server that dry-runs a prompt before an agent executes it.
Goal: estimate token consumption (input exact-ish; output as low/medium/high), score clarity and specification, and offer rewrites (reduce tokens / improve specification / both).
Evaluation and rewriting use Gemma 4 through the Gemini API (google-genai SDK; default model
env GEMMA_MODEL=gemma-4-26b-a4b-it; key in GEMINI_API_KEY).