Server has 9 tools with complete input schemas and descriptions, but quality is inconsistent. Tool names use German verbs (bundle_anlegen, karte_holen) which are non-standard for LLM-facing APIs, most production tools use English verb_noun patterns. Descriptions are present but brief (avg ~80 chars), lacking context on WHEN to use each tool or dependencies between them. No output schemas documented. Error handling is minimal, tools 'raise MCPFehler' but recovery guidance is absent. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels (WRITE, DESTRUCTIVE). Parameters lack format constraints and validation rules. No pagination support on list tools.

• Rename all tools to English verb_noun pattern: bundle_anlegen → create_bundle, bundle_holen → get_bundle, karte_holen → get_card, bundles_auflisten → list_bundles, karten_einfuegen → add_cards, karte_aendern → update_card, karte_loeschen → delete_card, bundle_aendern → update_bundle, bundle_anzeigen → display_bundle.
• Document output schemas for all tools. Example: get_bundle should return {slug, titel, beschreibung, gruppe, karten: [{id, art, vorderseite, rueckseite}], ...}. Include all fields downstream tools need.
• Add tool annotations: mark delete_card and create_bundle with destructiveHint=true; mark get_* tools with readOnlyHint=true; mark create_bundle with idempotentHint=true (if it validates before writing).
• Add pagination to list_bundles: accept limit (1-100, default 20) and offset parameters; return {bundles: [...], total_count, has_more}. Document the limit in the description.
• Enhance error messages with recovery guidance. Example: 'Bundle not found. Call list_bundles() to see available bundles.' Include valid enum values in parameter descriptions (e.g., art: 'flashcard|frage').