Loading ToolBench
Guarded golf retail workflow platform with deterministic orchestration, bounded model assistance, structured records, review queues, and operational audit trails. Exposes internal tools through an MCP-compatible REST adapter.
SwingOps exposes 11 tools with HTTP transport and reasonable naming conventions (verb_noun pattern: lookupProduct, findSimilarProducts, estimate, etc.). However, critical gaps undermine production readiness: (1) Most parameter descriptions are present but generic, many lack actionable constraints, format hints, or dependency documentation. (2) Output schemas are NOT documented in the provided source; tool responses are inferred but not formally specified. (3) Error handling is absent, no recovery guidance, no categorization of retryable vs. fatal errors. (4) Several tools lack required parameters (e.g., swingops.inventory.createSku requires only productId; no validation rules visible). (5) Tool descriptions are adequate (50 - 150 chars) but do not explain WHEN to use each tool vs. similar ones (e.g., lookupProduct vs. findSimilarProducts distinction is unclear). Baseline: 549 production tools average 194 chars for descriptions and 100% have documented return types; SwingOps shows neither. The server is functional but falls short of production-grade agent tool standards.
Send a customer message after approval. Disabled on the read-only connector surface.
Create an internal inventory SKU after approval. Disabled on the read-only connector surface.
Find similar internal product candidates when a parsed trade-in record is ambiguous.
Look up likely internal product and SKU matches for parsed golf trade-in equipment fields.
Dismiss a review queue item after a human reviewer determines no workflow update should be applied.
Look up one review queue item with its proposed structured golf-club data and workflow context.
Output schemas not documented. Tool responses are inferred but no formal return type specifications are visible in source. LLMs cannot plan downstream calls or extract required fields without documented output structure.
Error handling absent. No recovery guidance, error categorization, or actionable error messages visible. Agents cannot distinguish retryable failures from fatal errors or self-correct on invalid input.
Parameter descriptions lack actionable constraints. Many params (e.g., 'brand', 'productLine', 'conditionNotes') have generic descriptions without format hints, valid ranges, or dependency documentation. LLMs cannot validate input before calling.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | 2025-06-18+ | v2 |
List review queue items with workflow and intake context for human-in-the-loop workflow planning.
Resolve a review queue item after a human reviewer accepts or corrects the proposed workflow output.
Create a trade-in offer after approval. Disabled on the read-only connector surface.
Estimate a seeded demo trade-in range using parsed club fields, condition notes, accessory notes, and internal product match evidence.
Explain which condition and accessory factors changed a seeded demo trade-in valuation range.
Tool disambiguation unclear. lookupProduct vs. findSimilarProducts distinction is not explained. When should an agent call one vs. the other? Descriptions do not answer this, forcing LLM guesswork.
Write tools lack confirmation/dry-run pattern. createSku, create (tradeInOffer), resolve, dismiss, and send are destructive but show no confirmation step or preview capability. Agents can cause irreversible side effects without safeguards.