This MCP server has severe definition quality issues across all five tools. Tools are auto-generated from an OpenAPI spec (vision-agent-api), which explains systematic problems: (1) Descriptions are truncated/auto-generated placeholders like 'Tool Text To Od' (9 chars) and 'Tool Activity Recognition' (23 chars), below the 20-char hard minimum and far below the 50-200 char LLM-optimized baseline. (2) Parameter descriptions are almost entirely missing. The schema for text-to-object-detection includes 10+ parameters (timeout, Authorization, authorization, requestBody.image, requestBody.video, requestBody.prompts, requestBody.model, requestBody.nmsThreshold, requestBody.confidence, requestBody.jobId), none have actionable descriptions in the visible source. (3) Tool names do not follow verb_noun pattern; they use auto-generated API-style names like 'text-to-object-detection' instead of action-oriented names like 'detect_objects' or 'recognize_text'. (4) No parameter descriptions explaining valid ranges, formats, or dependencies. (5) No documented output schemas, the agent has no idea what fields to expect from these tools. (6) The `Authorization` and `authorization` parameters (duplicate?) expose authentication headers as tool parameters, violating secret-injection pattern. (7) requestBody parameter is a sprawling nested object without field-level guidance, unclear to LLMs which sub-fields are required or how to structure them. (8) No error handling or recovery guidance visible. These issues are consistent across all 5 tools, indicating a blanket auto-generation problem rather than isolated gaps.

• Rewrite tool descriptions to be 50 - 200 characters, LLM-optimized. For example: 'text-to-object-detection' → 'Detect and localize objects in images using natural language prompts. Returns bounding boxes, class labels, and confidence scores.' Include WHEN to use (vs other tools) and WHAT it returns.
• Add descriptions to every parameter. Examples: 'prompts: One or more text descriptions of objects to detect (e.g., "red car", "person"). At least one required.' 'model: The vision model to use. owlv2 is fastest, florence2 is most accurate. Default: owlv2.' 'confidence: Minimum confidence threshold (0 - 1). Detections below this are filtered. Default: 0.5.'
• Move authentication to server-side injection. Remove 'Authorization' and 'authorization' from tool parameters. Instead, load the API key from an environment variable (e.g., VISION_AGENT_API_KEY) in the index.ts main handler. This prevents credentials from leaking into traces.
• Rename tools to verb_noun convention: 'text-to-object-detection' → 'detect_objects_by_text', 'text-to-instance-segmentation' → 'segment_objects_by_text', 'activity-recognition' → 'recognize_activity', 'depth-pro' → 'estimate_depth', 'agentic-document-analysis' → 'analyze_document'.
• Document output schemas. Add a comment or code block in each tool's description explaining the response structure. E.g., 'Returns: {detections: [{bbox: [x1,y1,x2,y2], label: string, confidence: number}], metadata: {model: string, processing_time_ms: number}}'.
• Consolidate duplicate auth headers to a single, server-injected mechanism. Remove both 'Authorization' and 'authorization' parameters from tool schemas.