Cross-platform desktop automation with normalized coordinate protocol, precision-first perception, and local OCR
Server has 13 tools, but 8 are marked [HIDDEN] with empty or minimal schemas. Of the 5 visible tools, only 2 (computer_state, screen_interactor) have documented input schemas with properties and descriptions. The 3 visible tools with descriptions (computer_state, screen_interactor, skill, askuserquestion, ask_user_question) range from adequate to minimal. Hidden tools lack any visible schema or description detail, making them impossible to evaluate. No output schemas are documented. Error handling is basic (JSON-RPC errors only, no recovery guidance). Tool naming is reasonable (verb_noun pattern), but composition is poor: 8 hidden tools suggest internal implementation leakage rather than a clean public API. The server exposes low-level primitives (mouse, keyboard, screenshot, recorder) alongside high-level abstractions (computer_state, screen_interactor), violating single-responsibility principle.
Ask the user a question and wait for response (alias)
Ask the user a question and wait for response
[HIDDEN] Internal tool for executing computer actions
[HIDDEN] Internal tool for application management
Observe the current screen state with precision-first perception (UIA/AXAPI/AT-SPI2) or vision fallback (EasyOCR)
[HIDDEN] Internal tool for system information and profiling
[HIDDEN] Low-level keyboard control
8 of 13 tools are marked [HIDDEN] with no visible input schemas or descriptions. These tools are registered but not exposed to the public API, yet they appear in the tool registry. This violates transparency and makes evaluation impossible.
Hidden tools have empty or minimal input schemas (type:object with no properties). This prevents LLMs from understanding what parameters these tools accept.
screen_interactor accepts 'action' as a free-form string with no enum constraint. Valid actions (click, type, scroll, etc.) are mentioned in the description but not enforced in the schema. LLMs will hallucinate invalid action names.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 37 | 2026-07-28+ | v2 |
[HIDDEN] Low-level mouse control
[HIDDEN] Record screen activity
Execute actions on the screen (click, type, scroll) with coordinate normalization and precision targeting
[HIDDEN] Capture screenshot
Execute a named skill or macro
[HIDDEN] Internal tool for web navigation
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., what does computer_state return? What fields are in the response?). This violates pattern:tool-definition.
Duplicate tools: askuserquestion and ask_user_question are aliases with identical functionality. This violates composition principle (avoid multiple tools doing the same thing). LLMs waste reasoning cycles choosing between them.