cloudfact has 12 well-named tools with consistent verb-noun patterns (deploy, expose, list, catalog, etc.). All tools have descriptions (avg ~60 chars, within baseline). Input schemas are present and typed with Zod. Tool annotations (readOnlyHint, destructiveHint) are implemented. However, parameter descriptions are sparse or missing in several tools (e.g., 'deploy' has 8 params but only 3-4 have meaningful descriptions; 'expose' ssh param lacks detail). Output schemas are not documented. Error handling is basic (errorResult helper exists but no recovery guidance or categorization). No pagination support despite tools like 'catalog' and 'logs' that could return large result sets. Security: credentials tool accepts plaintext password param, should use server-side injection.

• Add substantive descriptions to all parameters. For 'deploy' backend param, explain: 'auto (default): auto-detect; tunnel: use Cloudflare Tunnel; workers: deploy to Workers.' For 'expose' ssh param, document expected structure: {host, user, port, key_path}.
• Document output schemas for all tools. E.g., 'catalog returns {project_name: string, deployments: [{name, url, status, created_at}]}'. Include this in tool descriptions or as a separate schema field.
• Remove password from credentials tool parameters. Instead, accept only username and note; store password via a separate secure channel (e.g., environment variable or vault lookup by deployment name).
• Add pagination to catalog and logs: accept limit (default 20, max 100) and offset/cursor parameters. Return total_count and next_cursor in response so LLM can iterate.
• Enhance error messages with recovery hints. E.g., 'Deployment "my-app" not found. Available: [list of names]. Call list() to refresh.' This enables LLM self-correction without extra round-trips.