Loading ToolBench
MCP server for AI-powered image editing and generation
12 tools with complete input schemas and descriptions, but significant quality gaps. Tool names lack action verbs (e.g., 'understand_image' vs 'analyze_image', 'clean_product_image' vs 'remove_background'). Descriptions are verbose (200-400 chars) with example values embedded, violating LLM-optimization guidelines. Parameter descriptions are present but generic. No output schemas documented. Error handling returns raw stack traces instead of recovery guidance. Three product-generation tools (necklace/bracelet variants) are near-duplicates with identical schemas, signaling composition issues.
Clean up product photos: remove backgrounds, watermarks, logos, and imperfections. Outputs professional white-background product images suitable for e-commerce. Accepts local file paths or URLs.
Clean product photos with pure white background: removes backgrounds, watermarks, logos, and imperfections. Outputs professional studio-quality images suitable for e-commerce. Accepts local file paths or URLs.
Edit images with text instructions: remove objects, change backgrounds, add elements, adjust colors, or apply effects. Uses Gemini 2.5 Flash for precise edits. Accepts local file paths or URLs.
Expand rectangular images to square format with AI-generated content. Uses Gemini 2.5 Flash to intelligently extend backgrounds and fill empty space. Outputs high-res WebP. Accepts local file paths or URLs.
Generate model wearing bracelet product photo from reference images. Creates professional lifestyle photography with model wearing the bracelet, suitable for e-commerce. Maintains exact product design from references while removing watermarks/logos.
Tool names lack action verbs. 'understand_image', 'clean_product_image', 'generate_necklace_white_background' do not start with clear verbs (analyze, remove, create). LLMs infer intent from names, weak names force reliance on descriptions and increase selection errors.
Six product-generation tools (necklace/bracelet × 3 styles) are near-duplicates with identical schemas and minimal description differences. This violates single-responsibility principle and forces LLMs to reason about subtle distinctions. Consolidate into parameterized variants (e.g., 'generate_product_photo' with product_type and style enums).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 46 | 2026-07-28+ | v2 |
Generate elegant tabletop product photo of a bracelet from reference images. Creates professional lifestyle photography with elegant styling, suitable for e-commerce. Maintains exact product design from references while removing watermarks/logos.
Generate premium white background product photo of a bracelet from reference images. Creates professional studio photography with pure white background, suitable for e-commerce. Maintains exact product design from references while removing watermarks/logos.
Generate photorealistic product images from text descriptions using Imagen AI. Supports 5 aspect ratios (1:1, 16:9, 9:16, 4:3, 3:4), outputs high-res WebP with AI quality review. Prompt structure: [Subject] + [Background] + [Style/modifiers] Example: 'Red ceramic mug, white background, studio lighting, macro lens, sharp focus, 4K' Modifiers: camera angles (aerial/45-degree/eye level), lighting (natural/studio/dramatic), focus (sharp/soft/depth of field), lenses (macro/wide-angle), quality enhancers (4K/photorealistic/professional).
Generate model wearing necklace product photo from reference images. Creates professional lifestyle photography with model wearing the necklace, suitable for e-commerce. Maintains exact product design from references while removing watermarks/logos.
Generate elegant tabletop product photo of a necklace from reference images. Creates professional lifestyle photography with elegant styling, suitable for e-commerce. Maintains exact product design from references while removing watermarks/logos.
Generate premium white background product photo of a necklace from reference images. Creates professional studio photography with pure white background, suitable for e-commerce. Maintains exact product design from references while removing watermarks/logos.
Ask questions about image content using Gemini 2.5 Pro vision. Extract text, identify objects, describe products, check quality, or get detailed descriptions. Accepts local file paths or URLs.
No output schemas documented. Tools return image paths/URLs but LLMs cannot infer response structure. Document: does generate_image return {url, path, base64, metadata}? What fields does understand_image return? Missing output schemas force LLMs to guess downstream field names.
Error handling returns raw stack traces (src/index.ts line 47: `error.stack`). LLMs cannot act on stack traces. Replace with actionable errors: 'Invalid image URL: must be http(s). Try a local file path instead.' or 'Gemini API rate limit exceeded. Retry in 60 seconds.'
Descriptions embed example values ('e.g., /path/to/image.jpg', 'e.g., https://example.com/image.jpg'). LLMs reuse examples literally in real calls, causing failures. Replace with format constraints: 'Local file path (absolute or relative) or HTTPS URL' and validate in code.