Loading ToolBench
MCP server for Gemini image generation (supports Nano Banana Pro & Nano Banana 2)
Three tools with well-structured schemas and generally good descriptions. All tools use proper JSON Schema with typed parameters and enum constraints. However, descriptions lack depth on use cases and recovery guidance. The schemas are complete and visible in the source code. Parameter descriptions are adequate but could be more instructive about valid ranges and relationships.
Analyze and describe one or more images using Google Gemini. Returns a text description of the image contents.
Edit one or more images using Google Gemini. Provide images and instructions for how to modify them. Returns a base64-encoded image.
Generate an image using Google Gemini. Optionally provide reference images to guide the generation style or content. Returns a base64-encoded image.
Tool descriptions lack recovery guidance and error context. Descriptions do not explain what happens on API failure, rate limiting, invalid images, or how to retry. Agents receive no actionable recovery paths.
No documented output schema for any tool. Tool descriptions say 'Returns a base64-encoded image' or 'Returns a text description' but do not describe the actual response object structure, fields, or types. LLMs cannot plan downstream tool calls or extract the right data.
generate_image and edit_image accept 'images' as optional/required reference input but descriptions do not explain limits (14 total, 10 object + 4 person refs). Parameter description states the limit but lacks clarity on when and how to use reference images.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 57 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 37 | - | v1 |
Descriptions do not warn about potential risks: generate_image and edit_image can create images of people (controlled by personGeneration enum). Descriptions should note this is subject to Gemini's usage policies and may be restricted. edit_image lacks warning that editing could produce unexpected output.
No error classification or guidance. If API key is invalid, rate limit is hit, image is too large, or Gemini rejects the request, the tool description does not explain whether the error is retryable, user-fixable, or fatal, or what the agent should do next.
outputPath parameter is optional in generate_image and edit_image. No description explains what happens if omitted, does the image remain in memory, get deleted, or returned inline? Agents need to know the lifecycle and whether saving is the default behavior.
thinkingConfig in generate_image is underdocumented. What does thinkingLevel do? When should agents use HIGH vs MINIMAL? What does includeThoughts do, does it add to the response? Does it slow down the call? Descriptions are too vague for LLM decision-making.
useGoogleSearch parameter in generate_image is boolean with minimal guidance. When would an agent enable this? Does it slow the call? Does it cost extra? Agents need to know the tradeoff to decide whether to use it.