DSH local image-processing MCP server (Pillow-based, offline, no API key). Exposes image tools for DeepSeek Harness via stdio transport.
12 image processing tools with complete input schemas and descriptions, but significant gaps in description quality, output schema documentation, and error handling guidance. Naming is clear and verb-driven (resize, crop, convert, etc.), and all tools follow a consistent pattern. However, descriptions are often generic or missing parameter-level detail. No documented output schemas beyond what is inferred from code. Error handling is minimal, most tools will fail silently or return stack traces rather than actionable recovery guidance. The server lacks tool annotations (readOnlyHint, destructiveHint) despite clear risk distinctions (info is READ_ONLY, others are WRITE). Parameter descriptions are minimal (e.g., 'Path to the image file' repeated 11 times). No per-parameter constraints (min/max, enum, pattern) for numeric or string inputs, inviting invalid LLM submissions.
Compress an image with quality control and format selection
Convert image to a different format
Crop an image to a rectangular region
Apply visual effects to an image (greyscale, sepia, invert, blur, sharpen, contrast, brightness)
Flip an image horizontally or vertically
Get image metadata (format, dimensions, mode, file size)
Overlay one image onto another at a specified position
Descriptions lack LLM-optimization depth and parameter-level clarity. Generic descriptions like 'Path to the image file' appear 11 times without guidance on format support, size limits, or expected encoding. Missing context on WHEN to use each tool vs. similar alternatives (resize vs. thumbnail, convert vs. compress, crop vs. overlay).
No documented output/return schemas. Code infers that tools return info_of(img, path) dictionary with 'path', 'format', 'width', 'height', 'mode', 'size_bytes', plus tool-specific fields (e.g., 'out', 'fit'). LLM cannot plan downstream calls or extract required fields without explicit schema documentation in the tool definition.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 54 | <=2025-11-25 | v2 |
Generate a placeholder image with optional text
Resize an image with optional fit mode (contain, cover, fill/stretch)
Rotate an image by a specified number of degrees
Create a thumbnail with maximum size constraint
Add text watermark to an image with customizable position, opacity, and color
Missing tool annotations despite clear semantic differences. 'info' is READ_ONLY; all others are WRITE/destructive. No readOnlyHint or destructiveHint declared in tool metadata. Agents cannot distinguish safe vs. risky operations without these hints.
No input constraints or ranges. Parameters like 'quality' (1 - 100), 'degrees' (any float), 'size' (pixel count), 'font_size' (positive int), 'opacity' (0.0 - 1.0), 'radius' (positive float) lack enum, minimum, maximum, or pattern constraints in schema. LLMs will pass invalid values (quality=999, degrees=NaN, opacity=2.5) that cause silent failures or exceptions.
Minimal error handling and no recovery guidance. Code contains try-except blocks that return raw exceptions (e.g., FileNotFoundError, ValueError). LLM receives stack traces or generic error dicts with no actionable next steps. Missing pattern guidance: 'File not found. Try checking the path or listing available images.' No categorization of errors (retryable vs. user-fixable).
Parameter interdependencies undocumented. 'effects' tool has conditional behavior: 'radius' applies only to blur; 'factor' applies only to contrast/brightness. No description states this dependency, so LLMs will pass radius when effect=greyscale (ignored silently) or omit radius when effect=blur (defaults to 2, which may not match intent).
Ambiguous parameter semantics and no natural identifier support. All tools require absolute file paths (strings). No support for relative paths, image IDs, or URLs. 'out' parameters are required and must be valid file paths, no inference or auto-naming offered. LLMs must reason about file system layout; natural interaction would be 'Save as thumbnail_200px.jpg in the same folder.'
Missing parameter descriptions for critical inputs. 'fit' parameter in resize: description says 'Fit mode: contain, cover, or fill/stretch (default: contain)' but does not explain WHEN to use each mode (aspect ratio preservation vs. stretching vs. filling). Similarly, position parameter in watermark lacks explanation of visual layout (where on screen each position places text).
No batch operations or efficiency guidance. Common workflows (e.g., batch resize 50 images, apply watermark to gallery) require 50 sequential tool calls. No batch_resize or bulk_watermark variants offered. Agents waste tokens and latency on loop constructs.