Loading ToolBench
MCP server for Gemini 3 integration with Claude Code - 30+ AI tools including image/video generation, deep research, code execution, and beautiful CLI
This server has 37 tools with basic descriptions and parameter schemas visible in tool-groups.ts. However, inspection reveals critical gaps in definition quality: (1) Descriptions are extremely brief (10-35 chars typically), well below the 50-200 char LLM-optimized baseline for A/B tier tools. (2) No visible input/output schemas in the source code provided, tool definitions appear to be metadata-only without formal JSON Schema. (3) Parameter types are declared (string, enum, object, array) but lack detailed constraints, ranges, or validation guidance. (4) Many tool names are generic action verbs (gemini-*) without clear verb_noun structure to disambiguate intent. (5) No visible error handling guidance, recovery paths, or confirmation steps for destructive operations (delete-cache, end-image-edit, run-code). (6) Critical security issue: gemini-run-code accepts arbitrary code execution with no dry-run or confirmation pattern. The server provides broad coverage (37 tools) but with shallow, formulaic documentation.
Analyze code for issues, improvements, and best practices
Analyze documents including PDFs and text files
Analyze an image and extract information
Analyze text for sentiment, key points, and structure
Analyze content from a URL
Generate creative ideas and brainstorming suggestions
Tool descriptions are critically short (avg 10-48 chars), well below LLM-optimized baseline of 50-200 chars. Examples: 'Query Gemini directly' (20 chars), 'List all active image editing sessions' (37 chars), 'Perform real-time web search using Google Search' (48 chars). These generic descriptions do not explain WHEN to use the tool vs similar ones, what it returns, or prerequisites.
No visible input/output schemas in provided source code. Tool definitions in tool-groups.ts declare only names, descriptions, and basic parameter enums (e.g. thinkingLevel for gemini-query). No explicit JSON Schema with proper type system, additionalProperties, required field arrays, or output documentation. Cannot verify schema completeness from static code analysis.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 54 | <=2025-11-25 | v2 |
| 2026-03-09 | D | 54 | - | v1 |
Check the status of a deep research task
Check the status of a video generation request
Compare content from multiple URLs
Continue an image editing session with a new instruction
Count the number of tokens in text
Create a cached prompt for faster repeated queries
Start a deep research task for comprehensive investigation
Delete a cached prompt
Create a multi-voice dialogue from text
End an image editing session and save the result
Extract structured data from text
Extract specific information from a URL
Extract tables from documents
Generate images using Gemini's Imagen model
Generate videos using Gemini's video generation model
Refine and improve image generation prompts
List all cached prompts
List all active image editing sessions
List all available TTS voices
Query Gemini directly
Query using a cached prompt
Send a follow-up question to an ongoing research task
Execute code and return the results
Perform real-time web search using Google Search
Convert text to speech with multiple voice options
Start an image editing session
Get structured output from Gemini with defined schema
Create a summary of text content
Create a summary of a PDF document
Analyze YouTube videos
Generate a summary of a YouTube video
Destructive tool gemini-run-code (IRREVERSIBLE risk) lacks confirmation, dry-run, or recovery guidance. Executing arbitrary code without a confirmation step or pre-execution review violates the confirmation-request pattern and creates unacceptable risk for agent-driven execution.
Destructive and WRITE-risk tools (gemini-delete-cache, gemini-generate-video, gemini-generate-image, gemini-speak, gemini-dialogue, gemini-deep-research, gemini-end-image-edit, gemini-continue-image-edit, gemini-start-image-edit) lack error handling guidance, recovery paths, or actionable error messages. LLMs cannot determine if a failure is retryable, user-fixable, or fatal.
Parameter descriptions are minimal or missing detail on constraints. Examples: gemini-query 'thinkingLevel' enum has no explanation of what each level means or when to use each; gemini-generate-image aspectRatio and imageSize enums lack guidance on use cases; gemini-analyze-code 'language' parameter lacks format constraints (e.g. valid language identifiers). Parameters require detailed format, range, and constraint documentation in descriptions.
No pagination or result limits documented for list/discovery tools. gemini-list-image-sessions, gemini-list-caches, gemini-list-voices have no indication of max results, pagination support, or offset/limit parameters. Large result sets risk context window exhaustion.
Stateful image editing (gemini-start-image-edit, gemini-continue-image-edit, gemini-end-image-edit) requires session management across tool calls but provides no guidance on session lifecycle, timeout, or cleanup. LLMs may not understand when to end sessions or handle session loss gracefully.
Generic tool naming with 'gemini-' prefix does not follow clear verb_noun convention. All 37 tools use 'gemini-<action>' format, which is provider-namespace rather than action-focused. Names like 'query', 'run-code', 'generate-image' are clear, but 'brainstorm', 'structured', 'extract' are ambiguous without context. Recommend: 'create_brainstorm', 'generate_structured_data', 'extract_data' for clarity.
Output schemas not documented. Tools like gemini-extract, gemini-structured, gemini-youtube-summary, gemini-deep-research, gemini-analyze-code return structured data but no descriptions of output fields, formats, or post-processing needs. LLMs cannot plan downstream tool chains without knowing what data is available.