Related Servers

Most tools map to clearly distinct actions (list, get, build_index, find_duplicates, quality_report), and the two search tools are separated by lengthy guidance on keyword vs. meaning vs. hybrid. The main friction is that calibre_search and calibre_semantic_search overlap in mode/scope combinations, and two tools (calibre_board_data, calibre_open_book) are explicitly marked 'do not call as the model', which is a selection hazard for an agent scanning the surface.

Every tool uses the same calibre_ prefix plus a snake_case verb_noun pattern (list_libraries, search, semantic_search, build_index, get_book, get_content, get_figures, list_categories, find_duplicates, quality_report, recover_metadata). The only outlier, calibre_ping, is a conventional health-check name and does not break predictability.

At 14 tools the set is well within a reasonable range for a library/search domain, and each tool covers a real capability (index build, search, extraction, figures, dedupe, quality). Two of the fourteen are internal widget endpoints that a model should never call, so the effective model-facing surface is slightly smaller than the count suggests.

The read/lookup surface is thorough, but the write half of the lifecycle is missing: there is no calibre_update_book or delete/add tool, and descriptions explicitly reference calibre_update_book and calibre_manage_bundles, neither of which exists in the set. calibre_recover_metadata therefore dead-ends — it returns a changes object meant for a tool the agent cannot call.