MCP server for vision-based image analysis, OCR, color analysis, and visual comparison using Ollama vision models and optional Tesseract OCR
VisionatorMCP provides 7 well-named vision analysis tools with comprehensive descriptions and detailed input schemas. Tool names follow verb_noun convention (vision_analyze, vision_color, vision_compare, etc.), making intent clear. Descriptions are substantial (150-400 chars) and explain modes, use cases, and parameters. Input schemas include types, descriptions, and defaults for most parameters. However, output schemas are not documented, responses are inferred from code but not formally specified in tool definitions. Error handling is present but generic (try/catch returning JSON errors without recovery guidance). Parameter descriptions could be more prescriptive about constraints (e.g., base64 encoding validation, file path validation). No tool annotations (readOnlyHint, destructiveHint) despite all tools being read-only operations.
Analyze, describe, or inspect an image using a vision language model. Consolidates multiple analysis modes into one tool: analyze (default, QA-focused), ask (free-form prompt), describe (crop to grid area), detect (automated QA scan for Unity visual issues). Optionally crop to a 3×2 grid area for focused analysis.
Analyze pixel colors in an image. Two modes: palette (default, extract dominant colors as ranked palette), match (calculate percentage of pixels matching custom color ranges). Both modes support sub-region analysis via relative coordinates (0.0–1.0). Works on original image data for accurate colors.
Compare two images/screenshots and identify visual differences between them. Returns structured analysis of what changed, where, and how significant the change is. Uses SkiaSharp for pixel-level comparison, augmented by VLM for semantic analysis. Image A is reference/before, Image B is actual/after.
Check the health and status of the Visionator MCP server. Verifies Ollama connection, lists available models (highlighting vision-capable ones), checks Tesseract availability, and reports current configuration. Use to diagnose connection issues or verify vision models are available.
Read an image file from the local filesystem and return it as a base64-encoded string ready for use with other vision_* tools. Convenience helper that eliminates manual base64-encoding. Image is automatically validated and optionally resized/compressed for optimal vision model performance.
Output schemas not documented. Tool descriptions explain what each tool does but do not specify the structure of returned JSON (fields, types, nested objects). LLMs cannot plan downstream operations or extract specific fields without knowing the response schema.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). All 7 tools are read-only vision analysis operations, but this is not declared in the tool metadata. Agents cannot infer safety properties without explicit annotations.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2026-07-28+ | v2 |
Extract or locate text in an image. Two modes: extract (default, OCR using VLM or Tesseract), find (locate specific text with bounding boxes). Supports VLM engine (more accurate for complex layouts) and Tesseract engine (faster, deterministic; falls back to VLM). Use for reading console errors, extracting UI labels, locating button text, verifying text placement.
Check if a described object, UI element, or visual feature is present in an image. Returns structured JSON with: whether object was found, confidence level, approximate location on screen, and visual description. Optionally crop to 3×2 grid area for region-specific search. Use for verifying UI buttons/panels exist, checking if game objects rendered, confirming error messages appeared, validating texture/asset loading.
Error handling lacks recovery guidance. Catch blocks return generic JSON errors (e.g., {error: ex.Message}) without actionable next steps. Per pattern:recovery-guide, errors should tell the LLM what to do: 'Invalid area. Try one of: top_left, top_center, ...' or 'File not found. Check the path and retry.'
Parameter constraints not fully specified. Several parameters accept free-form strings (e.g., 'mode', 'area', 'engine') with valid values documented in descriptions but not as enums in the schema. Per pattern:constrained-input, enums prevent hallucinated values and are self-documenting.
Mutually exclusive parameters not documented. vision_analyze accepts both image_b64 and image_path; vision_color accepts both image_b64 and image_path; vision_compare accepts both image_a_b64 and image_a_path. Descriptions say 'Use this INSTEAD of' but do not formally declare mutual exclusivity, risking LLM confusion.