Read-only MCP connector for GenieLocker private inference
Five read-only tools with clear, domain-specific descriptions (avg 150 chars). All tools have explicit Zod schemas with proper type constraints and descriptions. Tool annotations (readOnlyHint, idempotentHint) are correctly applied. Naming follows verb_noun pattern (genie_status, genie_catalog, genie_quote, genie_recipes, genie_credit_pricing). Schemas are well-formed with enums, min/max bounds, and defaults. Error handling returns structured JSON with isError flag. Main gaps: no output schema documentation, no pagination guidance for potentially large result sets (recipes), and parameter descriptions could be more prescriptive about valid formats.
Read only the lifecycle-verified SKUs that are currently eligible for purchase. Designed or benchmarked recipes are not represented as commercial inventory.
Read the current credit definition, burn multipliers, tier pricing, and examples. This never creates a wallet or buys credits.
Get a live, non-binding price quote. This is read-only and does not reserve capacity, create an account, or spend funds.
Search the public recipe menu by maturity, category, or text. The stage field is authoritative: only stage=commercial is live inventory.
Read current catalog, capacity, sellability, locker counts, and payment-rail status. This never creates or purchases a locker.
No output schema documentation. Tools return JSON via structuredContent but LLMs cannot infer response field names, types, or required fields. Agents must guess what fields to extract for downstream calls.
genie_recipes lacks pagination guidance. No limit parameter, no offset/cursor, no total count in response. If recipe catalog grows large, result could exceed context window.
Parameter descriptions lack format/constraint details. E.g., 'envelope' accepts 'l16' but no description of valid envelope types or how to discover them. 'placement' enum is ['resident'] but description doesn't explain why only one option exists.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 67 | 2025-06-18+ | v2 |
Error responses return raw Error.message without recovery guidance. E.g., network timeout or API 500 gives agent no actionable next step (retry? fallback? ask user?).
Tool titles are generic ('Check GenieLocker service status') and duplicate description content. Titles should be concise action phrases (e.g., 'Get service status') to aid LLM tool selection.