Loading ToolBench
A computer use agent skill that provides coordinate-based action planning for screen interactions
This server has critical deficiencies across all definition quality dimensions. Only 1 tool is defined (get_coordinates). The tool has a minimal description (8 words, well below the 10-1024 character baseline), no parameter descriptions, incomplete schema documentation, and no error handling guidance. The tool performs a non-trivial action (determining UI click coordinates) with only opaque 'screen_metadata' input and a freeform string output. No validation, no structured response schema, no recovery guidance on failure. The implementation is a stub method returning hardcoded mock data, not a real tool suitable for agent use.
Get next action coordinates based on screen metadata
Tool description is 8 characters ('Get next action coordinates based on screen metadata'), far below the 10-1024 character baseline and inadequate to guide LLM selection. Lacks WHAT, WHEN, and context.
Input parameter 'screen_metadata' has NO description. LLMs cannot infer what UI elements it should contain, what format is expected, or what happens if it's malformed. Required description is missing entirely.
Output schema is not documented. Tool returns a dict with string value {'next_action_coordinates': 'click x=450 y=320'} but LLMs have no way to know: What fields are in the response? What format are coordinates (x=N y=N or [x,y])? Can there be multiple actions? What does failure look like?
No error handling or recovery guidance. If screen_metadata is empty, malformed, or contains no detectable UI elements, tool returns generic output with no indication of failure state. LLM receives no actionable error message and cannot correct course.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 27 | 2026-07-28+ | v2 |
Input schema 'screen_metadata' is an untyped object with no property definitions. JSON Schema lacks 'properties' field. LLM cannot discover required fields, their types, or constraints. This violates the constrained-input pattern.
Tool performs a state-determining action (coordinates for computer use agent) but description does not explain: Is this read-only? Can it cause side effects? Is it idempotent? Risk marking (READ_ONLY) exists but tool description makes no mention of semantics.
Response is freeform string 'click x=450 y=320' not structured JSON with typed fields. LLM must parse the string to extract coordinates, error-prone and wastes tokens. Should return {action: 'click', x: 450, y: 320}.