Scored across 237 tools
Individual descriptions are unusually careful and frequently cross-reference siblings ('use X instead of Y'), which rescues a lot of borderline cases. But with 237 tools there are real overlapping clusters: list_todo / list_todo_for_course / get_todo_count / list_planner_items / list_upcoming_assignments / list_missing_submissions / list_overdue_assignments all answer 'what do I owe', and get_rubric_assessment vs get_my_rubric_assessment, list_quizzes vs list_new_quizzes, get_activity_stream vs get_course_activity_stream vs get_activity_stream_summary require reading the fine print to separate. An agent can usually get there, but misselection risk is meaningful.
The dominant convention is clear snake_case verb_noun (list_courses, get_assignment, submit_assignment_text, mark_topic_read), applied consistently across domains. Deviations are real but a minority: noun-only or phrase names like whoami, course_contacts, assignment_brief, daily_briefing, semester_snapshot, course_health_check, grade_impact_ranking, and smart_search. Readable and mostly predictable, just not uniform.
237 tools is far past any reasonable agent-facing surface and well into 'extreme mismatch' territory, especially since much of the bulk is redundant aggregation (daily_briefing, get_week_ahead, course_health_check, semester_snapshot, semester_wrap_up, diff_since_last_check, get_workload_forecast). Canvas is a large domain, but this set carries a lot of near-duplicate convenience wrappers alongside the core CRUD. An agent must spend significant context just discovering which of several equivalent entry points to use.
For a student-facing Canvas client the lifecycle coverage is excellent: courses, assignments, submissions, quizzes (both engines), discussions, announcements, inbox, calendar, files, modules, pages, outcomes, groups, grades and planner all have read plus the relevant student writes. Gaps are modest and mostly deliberate — no instructor/authoring operations (create assignment, grade, edit pages), no quiz taking (explicitly stated), no ePortfolio editing (a Canvas limitation), and no attachment support on Inbox messages. raw_canvas_request provides an escape hatch, so agents are unlikely to hit a hard dead end.