MCP server providing Gemini-powered web tools for search, page summarization, and image generation
Three tools with complete descriptions and clear naming. All tools have input schemas with typed parameters and descriptions. However, output schemas are undocumented (only inferred from code), error handling lacks recovery guidance, and no tool annotations present. Descriptions are solid (150-250 chars, well above 20-char minimum) but could be more concise for LLM efficiency. Parameter descriptions are present but some lack format/constraint details. No security concerns detected (no secrets in params). Composition is sound, each tool has one responsibility.
Generate, edit, or compose an image with Gemini's "Nano Banana" image model. Defaults to `gemini-3.1-flash-image-preview` (Nano Banana 2) — a native multimodal image model with strong prompt adherence, in-image text rendering, and up to 4K output. Every image carries an invisible SynthID watermark. The mode is driven by the prompt and the presence of `input_images`: - text only → generate from scratch - prompt + 1 image → edit, restyle, or transform that image - prompt + multiple images → blend, fuse, or use as subject/style references Be specific in the prompt: subject, style, composition, lighting, camera angle, and any literal text to render.
Summarize one or more web pages by URL using Gemini's URL Context tool. The model fetches each URL (HTML, PDF, JSON, plain text, or images up to 34 MB each, max 20 URLs per call) and returns a synthesized summary. With multiple URLs the model can compare, contrast, or consolidate across them — phrase `focus` accordingly (e.g. "diff the API changes" or "extract the pricing tier from each"). Public URLs only: no localhost, login-gated, paywalled pages, YouTube, Google Workspace docs, or other private content.
Search the web for current information, returning a synthesized answer with cited sources. Use this for any task that needs fresh web information — news, current events, latest releases, recent prices, "what's the latest...", "look up...", or any question whose answer might have changed since the model's training cutoff. Powered by Google Search via Gemini's `google_search` grounding, so coverage matches Google's index.
Output schemas not documented. web_search returns markdown with sources, summarize_pages returns markdown with URL metadata, generate_image returns a file path string, but none of these are formally specified in tool definitions. LLMs cannot plan downstream operations without knowing return structure.
No error handling guidance. Tools do not document what errors can occur (API failures, invalid URLs, rate limits, auth failures) or how to recover. LLMs receive raw exceptions with no actionable next steps.
Missing tool annotations. No readOnlyHint, destructiveHint, or idempotentHint declared. web_search and summarize_pages are read-only; generate_image is destructive (writes files). Annotations enable safer agent planning.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
generate_image parameter 'input_images' lacks format constraints in description. States 'absolute filesystem path or http(s):// URL' but does not specify max file size, supported MIME types, or max array length. Code enforces these but description does not guide LLM.
summarize_pages 'urls' parameter lacks validation guidance. Description does not mention max 20 URLs per call, unsupported URL types (localhost, paywalled, YouTube, Google Workspace), or max 34 MB per URL. LLMs may pass invalid URLs without warning.