Low-latency macOS GUI viewer MCP server — accessibility tree + targeted screenshots + navigation
jarvis-eyes-mcp has well-structured tool definitions with consistent naming (all start with 'eyes_' verb prefix) and comprehensive parameter schemas using Zod. Descriptions are present and moderately detailed (avg ~100 chars), explaining token costs and use cases. However, output schemas are not explicitly documented, responses are inferred from code (JSON text or base64 images). Error handling is minimal; most tools lack recovery guidance. Parameter descriptions are adequate but could be more prescriptive about constraints and formats. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications in metadata.
Arrange multiple windows in a layout (side-by-side, quad, or stack).
Click a UI element by label text, AX path, or 'x,y' coordinates.
Focus a window by app name or title.
Get the active window info: app name, title, bounds, focused element. Text only, ~20 tokens.
Read a specific UI element by accessibility path or role. ~20 tokens.
Read the menu bar items and keyboard shortcuts for the active app. ~100 tokens.
List all open tabs in the focused app (browser, terminal). ~30 tokens.
Output schemas not documented. Tools return JSON text or base64 images, but response structure is not formally specified. LLMs cannot reliably parse or chain results.
No tool annotations (readOnlyHint/destructiveHint/idempotentHint). Tools are classified as READ_ONLY or WRITE in metadata but not exposed via MCP tool annotations. Agents cannot infer safety properties.
Minimal error handling. Most tools return isError=true with a plain message (e.g. 'Element not found'). No recovery guidance, no actionable next steps for LLMs.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 69 | 2026-07-28+ | v2 |
Extract all visible text from active window. Returns plain text. ~50-100 tokens.
Read the accessibility tree of the active window. Returns hierarchical text. ~100-200 tokens vs 3000+ for a screenshot.
Send a keyboard shortcut (e.g., 'cmd+s', 'ctrl+shift+n', 'escape').
List all visible windows with app name, title, bounds. ~50 tokens.
Capture just a specific UI element by accessibility path. Smallest possible image.
Capture a specific screen region as PNG. Use ONLY for charts/images that text cannot describe.
Scroll in a direction (up/down/left/right) by N clicks.
Type text into the focused element.
Parameter descriptions lack format/constraint details. E.g., 'path' in eyes_get_element says 'Accessibility path like AXWindow/AXToolbar/AXButton[Save]' but does not specify required format, valid characters, or error cases.
Cache logic in eyes_get_ui_tree is confusing. Condition 'if (use_cache !== false)' then calls getUITree() regardless, making cache check ineffective. Cached state is stored but never returned.