Loading ToolBench
MCP server for LoRA discovery and image generation using Flux
Server has 3 tools with explicit schemas visible in src/mcp/server.ts and type definitions. Naming follows verb_noun pattern (discover_, generate_, save_), which is good. Descriptions are present but vary in quality. Critical gaps: discover_loras has query/style/limit with minimal parameter descriptions; generate_image has complex nested schema for loras array but loras parameter description is sparse; save_image_to_disk has only 'url' parameter with a generic description. No error recovery guidance in descriptions. Output schemas are documented in code but descriptions lack guidance on what LLM should do if errors occur. Tool annotations present (destructiveHint, streamable, title) which is positive. Overall falls into 'fair' territory with noticeable gaps in parameter documentation and error handling patterns.
Searches for LoRA models based on query, style, or other filters.
Generates an image based on a prompt and optional parameters.
Saves an image from a base64 data string to the local filesystem.
save_image_to_disk parameter 'url' has only generic description ('URL or base64 data string of the image to save'). Lacks clarity on expected format, length constraints, or what constitutes valid input. LLM cannot determine whether to pass raw base64 or data: URI.
generate_image loras parameter is complex nested array of objects with 'path' and 'scale' properties, but the parameter description only says 'Array of LoRA models to apply' with no guidance on valid path format, scale range (0-1? 0-10?), or examples. Nested schema properties lack individual descriptions.
discover_loras parameters (query, style, limit) have descriptions but lack constraints. Is 'style' an enum or free-text? What are valid values? Limit is listed as number with default 20 but no min/max bounds stated. LLM cannot determine if passing limit=1000 is acceptable.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 51 | 2025-06-18+ | v2 |
| 2026-03-09 | F | 42 | - | v1 |
No error recovery guidance in any tool description. If discover_loras returns empty results, the LLM has no hint to try different search terms or check if the service is available. If save_image_to_disk fails, no guidance on whether to retry or ask user for a different path.
generate_image has guidance_scale parameter (float, default 7.5) but no stated range. Is 1-20 valid? 0-1? Can it be negative? This invites LLM to pass arbitrary values that may break the API or produce nonsensical results.