MCP server that drives the local desktop with computer use actions (click, type, scroll, screenshot, etc.). Supports batched mode (OpenAI-style actions[]) or split mode (per-action tools).
This server exposes 10 computer-use tools with complete JSON schemas and basic descriptions. However, descriptions are minimal (10-40 chars), parameter descriptions are absent, and error handling guidance is missing. The batched `computer_use` tool has a complex oneOf schema but no guidance on action ordering, failure modes, or recovery. Split-mode tools are thin wrappers with identical minimal descriptions. No parameter-level documentation explains constraints, formats, or dependencies. Output schemas are not documented. The server lacks actionable error messages and confirmation patterns for irreversible operations (click, drag, double_click marked IRREVERSIBLE but no safeguards).
Computer use: click
Computer use: double_click
Computer use: drag
Computer use: keypress
Computer use: move
Computer use: screenshot
Computer use: scroll
Computer use: type
Parameter descriptions missing entirely. Tools accept x, y, scroll_x, scroll_y, keys, button, text, path but provide no guidance on valid ranges, formats, or constraints. LLMs cannot infer whether x=0 is valid, whether keys=['ctrl','a'] is the right format, or what button values mean.
Tool descriptions are 10 - 40 characters, below the 50-char minimum for LLM selection. 'Computer use: click' and 'Computer use: type' do not explain when to use each tool, what happens, or how they differ from similar tools. Descriptions should be 50 - 200 chars with context.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 38 | <=2025-11-25 | v2 |
Execute OpenAI-style computer use actions in order (batched).
Computer use: wait
No error handling or recovery guidance. Tools marked IRREVERSIBLE (click, drag, double_click) have no confirmation pattern, dry-run option, or error messages explaining what went wrong or how to retry. Agents cannot distinguish retryable failures from permanent errors.
Output schema not documented. Tools return JSON via response_to_json() but no schema describes the response structure, fields, or types. LLMs cannot plan downstream actions or extract data reliably.
Batched computer_use tool has complex oneOf schema but no guidance on action ordering, partial failures, or rollback. If action 3 of 5 fails, what happens to actions 4 - 5? Can the agent retry just action 3? No documentation.