ClawGuard exposes 9 tools across HTTP API and CLI. Most tools have descriptions (7/9), but parameter documentation is sparse and inconsistent. Input schemas are present but lack depth: many parameters missing type constraints, enums, or validation guidance. Output schemas are partially documented in response models but not exposed in tool definitions. Error handling is basic (HTTP exceptions) without recovery guidance. No tool annotations (readOnlyHint/destructiveHint). The 'scan' and 'scan_all' CLI tools appear inferred rather than explicitly registered, capping their scores at 50. Overall, definitions lack the LLM-optimization and constraint clarity needed for reliable agent invocation.

• Explicitly register all 9 tools in the MCP server definition with full JSON Schema input/output. CLI tools must be wrapped as MCP tools, not inferred.
• Add parameter descriptions for all inputs. Example: 'url: ClawHub skill URL (https://clawhub.ai/skills/...). Either url or path required.' 'no_llm: Skip LLM-based analysis for faster, free scanning (uses only static analyzers).'
• Add enum constraints and bounds. 'limit' in list_scans: integer, minimum 1, maximum 100, default 20. 'json_output': boolean, default false.
• Enhance error responses with recovery guidance. Instead of 'Scan failed: <error>', return '{"error": "Scan failed", "detail": "<specific reason>", "recovery": "Try uploading a valid SKILL.md archive or check the URL format.", "retryable": true}'.
• Add tool annotations. Mark submit_scan, upload_scan, scan, scan_all, bulk_scan with destructiveHint=true (they modify state). Mark health, get_scan, list_scans, watch with readOnlyHint=true.