Loading ToolBench
Medical calculator server providing clinical scoring tools and risk assessment instruments for emergency medicine, geriatrics, gastroenterology, gynecology, and domestic abuse screening.
Server provides 7 medical/clinical assessment tools with generally well-structured Pydantic models and detailed input schemas. All tools have clear descriptions (100-300+ characters) explaining clinical purpose and decision-making context. Input parameters are typed with enums and constraints. However, output schemas are minimally documented in the source code shown, and the tool names don't follow action-verb conventions typical of agent tools (e.g., 'acep_ed_covid19_management_tool' reads as a noun phrase, not 'calculate_acep_ed_covid19_disposition'). Tools are read-only clinical calculators, no destructive operations, which reduces error-handling complexity but also means no confirmation/dry-run patterns are needed. Composite output structures are defined but lack inline field descriptions in responses. No tool annotations (readOnlyHint, etc.) visible in schema definitions.
Determines recommended disposition for adults with suspected or confirmed SARS-CoV-2 infection in the emergency department using NIH severity classification integrated with PRIEST score and patient risk factors to guide diagnostic workup and management.
Diagnostic medical tool designed to classify the form of human faeces into seven categories to evaluate bowel health, diagnose digestive issues, and monitor treatment effectiveness based on stool consistency and intestinal transit time.
Evidence-based counseling tool developed at Brigham and Women's Hospital that predicts the probability of achieving at least one, two, or three live births based on a woman's age at egg freezing and the number of mature eggs retrieved.
Validated lethality risk assessment instrument designed to predict the risk of intimate partner femicide using 20 risk factors based on adjusted odds ratios from multisite case-control study of 310 femicide cases.
Validated instrument that identifies older cancer patients who could benefit from comprehensive geriatric assessment (CGA) by incorporating 7 items from the Mini Nutritional Assessment (MNA) plus patient age to screen for geriatric vulnerability.
Tool names do not follow action-verb conventions. Names like 'acep_ed_covid19_management_tool' and 'woman_abuse_screening_tool' are noun phrases describing the tool itself, not the action it performs. LLMs infer action intent from verb-first names (e.g., 'calculate_acep_ed_covid19_disposition', 'screen_woman_abuse_risk'). This increases agent confusion when selecting among similar clinical tools.
Output schemas (AcepEdCovid19ManagementToolResponse, etc.) define response models with fields like 'result', 'unit', 'interpretation', 'stage', 'stage_description' but lack inline field descriptions within the response schemas themselves. The Config.json_schema_extra provides only an example, not field-level documentation. LLMs cannot determine what each output field means without explicit descriptions.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 61 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 43 | - | v1 |
Validated screening tool developed by BAPEN to identify adults who are malnourished or at risk of malnutrition using three key indicators: BMI, recent weight loss, and acute disease effects.
Validated 8-question screening instrument designed to detect intimate partner violence (domestic abuse) in healthcare settings by assessing relationship dynamics, emotional impact, and history of physical, emotional, and sexual abuse.
No tool annotations visible in schema definitions. Tools should declare readOnlyHint=true (all tools are read-only calculators with no side effects). This metadata helps agents understand tool safety and prevents unnecessary confirmation logic. Current spec supports tool annotations in input/output schemas.
No explicit error handling documentation visible in tool definitions. Clinical tools should document expected error scenarios (e.g., invalid clinical parameters, out-of-range values) and guide LLM recovery. Current implementation shows validation (via Pydantic min/max/enum) but no error response schema or recovery guidance.
Some enum descriptions in danger_assessment_tool contain example field names (e.g., 'controls_activities' field description says 'Have you ever been beaten by him when he was drunk or high on drugs?' which should be 'controls_daily_activities'). This mismatch between parameter name and description text will confuse LLMs during parameter binding.