Related Servers

Most tools target distinct resource+action combos, but there are genuine overlaps: can_i_ship and get_eval_gate both serve as CI-gate verdicts on finished eval runs, and is_my_judge_trustworthy largely re-reads trust/TPR/TNR/drift data already present in get_criterion and list_criteria. Additionally, create_eval and screen_my_traffic both support 'is a cheaper model better' screening, though the descriptions try to disambiguate them.

The dominant pattern is clean verb_noun snake_case (list_*, create_*, get_*, update_*, delete_*, cancel_*) applied consistently across 20+ resource domains. The exceptions are notable but few: the first-person sentence-style names can_i_ship, is_my_judge_trustworthy, and screen_my_traffic break the convention, and upsert_alias/auto_improve_criterion deviate slightly from standard CRUD verbs.

81 tools is far beyond a typical MCP surface and will substantially bloat agent context and increase misselection risk, even though the underlying platform genuinely spans many domains (evals, fine-tuning, GRPO, dedicated endpoints, audit, labels, datasets). The breadth is real but the tool count is still excessive for an agent-facing interface, sitting at the extreme end of 'too many.'

Core lifecycles are thoroughly covered for most resources: criteria, evals, dedicated endpoints, batches, fine-tuning jobs, GRPO runs, model versions, and logs all have create/read/list/update/delete or cancel where appropriate. Notable gaps exist: datasets are create-only (no list/get/delete for created datasets), label sets cannot be deleted, training files cannot be removed, and labels have no delete operation—gaps that will matter for cleanup workflows in an audit-focused platform.