MCP server for inspecting and interacting with JavaFX scene graphs in running applications
This JavaFX inspection MCP server presents a cohesive tool set with strong naming conventions and reasonable descriptions. All 14 tools follow verb_noun patterns (get*, set*, connect*, disconnect*, etc.) which is excellent for LLM parsing. However, there are critical gaps in parameter documentation, output schema definition, and error handling guidance. Most tools lack descriptions of what they return and how to handle failures. Parameter descriptions exist but many lack format constraints, valid ranges, or dependency hints. No evidence of output schema documentation. Tools are well-composed (each handles one concern) and input schemas are visible and typed, but incomplete, many lack required descriptions for all parameters or provide only minimal guidance. The server targets a specialized domain (JavaFX UI inspection) with clear intent, but production readiness is undermined by missing output documentation and sparse error guidance.
Fire a ButtonBase action without mouse input
Save a silent MP4/H.264 clip of a node or window scene
Perform a click action on a node using synthetic gesture or robot
Connect to a running JavaFX application by PID
Disconnect from a running JavaFX application
Discover running JavaFX applications
Get detailed information about a specific node
No output schema documentation for any tool. Tools return complex objects (scene graphs, node details, stage lists, video/screenshot paths) but LLMs have no documented structure to parse responses or chain results. This forces LLMs to infer output format from trial and error.
Parameter 'value' in setProperty lacks type definition and description of what formats are accepted. Schema shows description exists but constraint information is vague ('Value to set (type depends on property)'). LLMs cannot infer whether to pass a string, number, object, or array without concrete examples or enum definitions.
Tools dealing with multi-step inspection (connectApplication → getStages → getScenegraph) lack dependency documentation. Descriptions do not indicate 'you must call connectApplication before calling getStages', forcing LLMs to discover constraints through failure.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Get the scene graph tree for a window in a connected application
Get list of windows (Stages and PopupWindows) in a connected application
Request keyboard focus for a node
Highlight a node with a red border overlay
Set a property on a node
Save a PNG screenshot of a node or window scene
Type a key into a node or the focused node
No error recovery guidance in any tool description. How should an LLM handle 'Application not found'? 'PID is invalid'? 'Node does not support this property'? Raw errors provide no next steps, violating recovery-guide pattern.
Parameters like 'depth', 'maxWidth', 'maxHeight', 'durationSeconds', 'framesPerSecond' lack explicit range constraints in descriptions (e.g., 'depth: 1-50', 'framesPerSecond: 1-30 fps'). Schema shows some constraints but descriptions do not articulate them in natural language for LLM parsing.
Tools like clickNode, activateNode, typeKey modify application state (WRITE risk) but descriptions do not explicitly state 'This modifies the target application' or warn about side effects. Agents need to know which calls are safe to retry.
propertyFilter and mode parameters in several tools accept arrays or enums but descriptions do not enumerate allowed values or provide examples. LLMs cannot determine valid property names or click modes without a reference list.