Three well-named, read-only tools with clear descriptions and complete input schemas. All tools follow verb_noun naming (search_, query_, layer_info). Descriptions are concise (100 - 200 chars) and explain WHAT, WHEN, and WHAT-TO-DO-NEXT. Input schemas include types and descriptions for all parameters. However, output schemas are not documented, LLMs cannot predict response structure for downstream chaining. Parameter descriptions lack explicit constraints (e.g., 'limit: 1 - 50' is stated but not in the description text itself). No error handling guidance or recovery paths documented. Security is solid (SSRF guard in layerUrl, no secrets exposed), but no permission gates or audit trail patterns.

• Document output schemas for all three tools. Example for search_datasets: 'Returns {total: number|null, count: number, datasets: [{id, name, summary, org, org_id, record_count, type, url}]}'. This enables LLMs to extract dataset URLs for downstream query_layer calls.
• Add error recovery guidance. Replace 'throw new Error("ArcGIS Hub: 404")' with structured error responses: '{error: "dataset_not_found", message: "No datasets match your query. Try broader keywords like \"parcels\" or \"zoning\".", retryable: false}'. Guide LLMs on next steps.
• Embed constraint descriptions in parameter definitions. Change 'Max datasets (1-50, default 20)' to 'Max datasets to return (1 - 50, default 20). Larger limits may slow queries.' Make constraints machine-readable via JSON Schema minItems/maxItems.
• Add pagination support to search_datasets. Return {total, count, datasets, next_page_token} and accept an optional page_token parameter. Document: 'If total > count, call again with page_token to fetch the next batch.'
• Validate SQL where clauses in query_layer. Reject clauses containing DROP, DELETE, INSERT, or other DDL keywords. Return: 'Invalid WHERE clause: DDL not allowed. Use SELECT-safe conditions only.' This prevents accidental or malicious data modification.