The server exposes a single tool 'compress_prompt' with a clear verb-noun naming convention. The tool description is comprehensive (180+ characters) and explains what it does, when to use it, and dependencies. However, the input schema is minimal (single 'text' parameter with basic type and description), and there is NO documented output schema, the response is unstructured text with embedded metadata rather than a typed object. The error handling returns the original text gracefully but does not guide the LLM on recovery paths or categorize errors. The server lacks tool annotations (readOnlyHint, destructiveHint, idempotentHint) and does not follow structured output patterns used by A-grade tools. The implementation is functional but falls short of production-grade quality standards for agent-facing tools.

• Define a structured output schema for compress_prompt. Instead of returning unstructured text, return a JSON object: {original_text, compressed_text, mode, coverage_percent, tokens_in, tokens_out, tokens_saved}. This allows agents to extract fields reliably and chain results into downstream tools.
• Add constraints to the 'text' input parameter: specify maximum length (e.g., 'up to 10,000 characters'), expected encoding (UTF-8), and what happens if input exceeds limits. Example: 'The prompt text to compress (UTF-8, max 10,000 characters). Texts exceeding the limit will be truncated or rejected.'
• Add tool annotations. In the server.py tool registration, add readOnlyHint=true and idempotentHint=true to signal that this tool does not modify state and can safely be retried.
• Improve error handling. Catch specific exceptions (Ollama connection errors, model loading errors, validation failures) and return categorized errors: {error_type: 'ollama_unavailable'|'model_missing'|'validation_failed', message: '<actionable recovery guidance>', recovery_steps: [...]}. Example: 'Ollama is not running on localhost:11434. Start Ollama and pull the required models: ollama pull llama3.2:1b && ollama pull nomic-embed-text.'
• Add a startup health check. In main(), before entering the STDIO loop, verify that Ollama is reachable and both required models are loaded. Log warnings if models are missing, and reject tool calls with clear error messages rather than failing silently.
• Document parameter validation rules in the 'text' parameter description. Include examples of what compression preserves (conditionals, negations, critical intent) and what it may lose (verbose explanations, redundant examples). This helps agents understand the trade-off.