Loading ToolBench
MCP server for Condor Eye — gives Claude agents the ability to see the screen
Condor Eye MCP has well-structured tool definitions with complete input schemas and clear descriptions. All 8 tools are explicitly registered with proper JSON Schema. Naming follows verb_noun convention (condor_eye_capture, condor_audio_start). Descriptions are substantive (avg ~150 chars) and explain WHEN to use each tool. However, output schemas are entirely undocumented, LLMs cannot predict response structure. Parameter descriptions are present but lack format constraints (e.g., region coordinates have no min/max bounds). Error handling is absent from tool definitions. No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles (condor_audio_start is WRITE, others are READ_ONLY). Missing idempotency guidance for retry-safe operations.
Fetch the latest transcript text for an active tap.
Start a Condor Audio tap for a target app.
Check the Condor Audio API status, configured targets, and active taps.
Stop an active audio tap by tap id.
Capture a screen region and get an AI description of what's visible. If no region is specified, captures the full screen. Returns a text description of the screen content.
Find a UI element on screen using AI vision. Captures the full screen and uses AI to identify the target and return its bounding box. Prefer condor_eye_windows for finding windows by title (free and instant) — use this tool for finding specific UI elements within a window.
No output schemas documented. LLMs cannot predict response structure (fields, types, pagination). Forces agents to guess what data is available after each call.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). condor_audio_start and condor_audio_stop are clearly WRITE operations but lack destructiveHint. Agents cannot distinguish safe from risky calls.
Numeric parameters lack bounds. 'region' coordinates (x, y, width, height) have no min/max constraints. LLMs may pass negative or absurdly large values. Add constraints like 'x: integer, minimum: 0, maximum: 65535'.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
Check if the Condor Eye app is running and get its current configuration.
List visible windows with their screen positions and PIDs. Use this to find a window's exact bounds before capturing it — avoids expensive AI-based locate.
No error handling guidance in tool definitions. If Condor Eye app is unreachable or audio tap fails, agents have no recovery path. Add error classification and retry guidance to descriptions.
condor_eye_status and condor_audio_status descriptions are generic (60 chars). Do not explain what 'configuration' or 'active taps' means. Expand to clarify: 'Returns API version, model name, and list of currently active audio taps with their app targets and PIDs.'