Loading ToolBench
MCP server exposing Reticle's VLM-based screen-region grounding to MCP-compatible AI agents. Uses Qwen2.5-VL on Apple Silicon via mlx-swift-lm to detect bounding-box regions in images.
Single tool 'ground_region' has a well-structured schema with proper JSON Schema format and clear parameter descriptions. Tool description is comprehensive (280+ chars) and explains the VLM capability, model constraints, and performance characteristics. However, output schema is not formally documented, the tool returns JSON-encoded bounding boxes but the response structure is only described in prose. Parameter descriptions are detailed and actionable. No tool annotations (readOnlyHint, idempotentHint) are present. Error handling is basic but functional.
Detect bounding-box regions in an image using a vision-language model (Qwen2.5-VL on Apple Silicon, via mlx-swift-lm). Pass a path to a local PNG/JPEG/TIFF/HEIF and a natural-language prompt describing what to find. Returns model-coordinate bounding boxes (max 1280 px on the longest side, snapped to multiples of 28 — the patch grid Qwen2.5-VL was trained on). First call loads the model (~25 s); subsequent calls are fast.
Output schema not formally documented. Tool returns JSON array of bounding boxes but response structure (fields: label, x1, y1, x2, y2) is only described in prose, not in a structured schema definition. LLMs cannot reliably parse or chain this output.
Tool name 'ground_region' lacks action verb prefix. Follows pattern 'noun_verb' rather than 'verb_noun'. Should be 'detect_regions' or 'ground_image_regions' to match LLM naming conventions (get_, create_, search_, detect_).
No tool annotations present. Tool is read-only (does not modify state) but lacks readOnlyHint annotation. This prevents clients from optimizing caching or parallelization strategies.
Error handling returns plain text error messages without structured guidance. Errors like 'File not found' or 'Could not decode image' do not suggest recovery steps (e.g., 'Check file path' or 'Supported formats: PNG, JPEG, TIFF, HEIF').
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Parameter 'prompt' description includes example prompt text but does not formally constrain format or length. LLMs may generate arbitrarily long or malformed prompts without validation guidance.