MCP server wrapping Codex Computer Use for Claude Code CLI on Windows
Server defines 13 tools with explicit schemas and descriptions. However, descriptions are minimal (10-50 chars), parameter descriptions are sparse or missing, and output schemas are undocumented. No error recovery guidance. Tool names follow verb_noun convention (list_, get_, activate_, click, etc.), which is good. Schemas are present but lack depth: many parameters have no description, and complex nested objects (window) lack field-level documentation. No tool annotations (readOnlyHint, destructiveHint). Error handling returns generic text results without actionable recovery paths. Composition is reasonable, tools are single-purpose, but output chaining is unclear (e.g., does list_windows return window objects that activate_window accepts?).
Bring a window to the foreground
Click at coordinates or on an element
Drag from one position to another in a window
Rehydrate a window object by id (useful after losing window binding)
Capture screenshot and/or accessibility tree of a window. Call this before input actions.
Launch an application
List installed apps with their open windows
Tool descriptions are too brief (10-50 chars). Rubric baseline is 194 chars. Descriptions lack WHEN to use, WHAT it returns, and prerequisites. Example: 'List all open windows that can be controlled' does not explain what fields are returned or how to use the result with activate_window.
Parameter descriptions are missing or minimal. The 'window' parameter (used in 8 tools) has no description of its structure or required fields. The 'element_index' parameter lacks guidance on how to obtain it. Rubric requires every parameter to have a non-empty description.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 49 | 2026-07-28+ | v2 |
List all open windows that can be controlled
Perform a secondary action on an element
Press a key in a window
Scroll in a window
Set the value of an element
Type text into a window
Output schemas are not documented. Tools return ToolResult with generic 'text' or 'json' content, but the structure of returned data (e.g., what fields does list_windows return? what is the schema of get_window_state output?) is not specified. LLMs cannot plan downstream calls without knowing output structure.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools like 'click', 'type_text', 'drag' are clearly destructive/write operations, but the schema does not declare this. Agents cannot reason about safety or retry semantics without annotations.
Error handling returns generic text results ('Clicked', 'Typed: ...') without recovery guidance. If a click fails, the LLM receives no actionable error message. Rubric requires errors to tell the agent what to do next and categorize as retryable/user-fixable/fatal.