Loading ToolBench
Multi-AI MCP bridge: Gemini + OpenRouter (400+ models) — text, code, image, video, TTS, RAG, Deep Research
The server provides 21 well-named tools covering diverse AI capabilities (text, image, video, audio, RAG, research). Most tools have clear descriptions (50-200 chars range) and properly typed input schemas with enums for constrained parameters. However, critical gaps exist: (1) NO output schemas are documented for any tool, the rubric requires structured response documentation; (2) Error handling is generic/absent, tools don't guide recovery ('retryable vs user-fixable'); (3) Several tools accept file paths without validation guidance; (4) Some parameter descriptions lack format/constraint detail; (5) Destructive tools (delete_conversation) lack confirmation patterns. The average tool scores 61 across naming (95), description (70), and schema (40) dimensions. Naming is excellent (verb_noun convention consistently applied), but schema/output documentation is the primary drag.
Ask any model (Gemini or OpenRouter: 400+ models). Auto-detects model type and routes request accordingly.
Analyze large codebases using Gemini's 1M token context window. Perfect for architecture analysis, cross-file review, and understanding complex projects.
Analyze images using Gemini vision capabilities. Describe, extract text (OCR), identify objects, or answer questions about images.
Ask Gemini any question. Get detailed responses, code explanations, creative writing, brainstorming, or analysis.
Brainstorm ideas using structured frameworks (divergent, convergent, SCAMPER, design thinking, lateral thinking).
Critique and challenge an idea, plan, or code using Gemini to identify weaknesses, risks, and improvements.
No output schemas documented for any tool. Rubric requires documented return types (100% of A+ tools). Without output schemas, LLMs cannot plan downstream tool chains or extract expected fields. Example: gemini_analyze_codebase returns 'str' but users need to know if it's structured JSON, markdown, or plain text, and what fields are included.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 60 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 44 | - | v1 |
Review code for quality, security, performance, readability, and bugs using Gemini.
Create a file store for RAG with File Search.
Execute comprehensive research using Google's Deep Research Agent. Autonomously conducts multi-step web searches and synthesizes findings. The agent plans and executes research strategy, conducts multiple targeted web searches, synthesizes findings into comprehensive reports, and provides citations and sources.
Delete a saved conversation by ID.
Search across files in a file store using semantic search powered by Gemini embeddings and retrieval.
Generate code in any language using Gemini with optional context from your files.
Generate images using Gemini/Imagen native image generation. Use descriptive prompts (not keywords) for best results.
Generate videos using Google Veo 3.1 with native audio. Creates 4-8 second videos with realistic motion, dialogue, and sound effects.
List saved conversations from local SQLite storage or cloud Interactions API.
List all file stores.
List all available Gemini models (text, image, video, TTS, speech-to-text).
Convert text to speech with 30 voice options. Supports single and multi-speaker (up to 2) audio.
Transcribe audio files (WAV, MP3, FLAC, OGG, AIFF, ULAW) to text using Gemini's speech-to-text model.
Upload a file to a file store for RAG indexing.
Search the web using Gemini with Google Search grounding. Returns answers with citations from authoritative sources.
No error handling guidance in any tool. Rubric requires: retryable/user-fixable/fatal classification and actionable recovery hints (e.g., 'Try search_users() first'). Example: gemini_delete_conversation does not explain what happens if conversation_id is invalid, or if deletion is retryable.
Destructive tool (gemini_delete_conversation) lacks confirmation pattern. Rubric pattern:confirmation-request states irreversible operations should support dry-run or confirmation step to prevent agent mistakes.
File path parameters (gemini_analyze_codebase 'files', gemini_analyze_image 'image_path', gemini_upload_file 'file_path') lack validation guidance. Descriptions do not specify allowed formats, path traversal protections, or what happens if files don't exist. Rubric pattern:tool-gateway requires sanitization guidance and format specs.
gemini_text_to_speech 'speakers' parameter is defined as array of objects but lacks schema detail. Description says 'Multi-speaker configuration (up to 2)' but doesn't specify required fields, format, or allowed values for speaker objects.
Several tools (gemini_brainstorm, gemini_challenge) have optional parameters without clear defaults or consequences. Example: gemini_brainstorm 'constraints' is optional, unclear if omitting it means 'no constraints' or 'use defaults'. Rubric requires documenting defaults and their consequences.
List tools (gemini_list_conversations, gemini_list_models, gemini_list_file_stores) return no parameters. Rubric pattern:paginated-result requires pagination support (limit, offset/cursor) and total counts for list endpoints. Without pagination, large result sets blow context.
ask_model parameter 'model' accepts string without constraint/enum (e.g., 'openai/gpt-4o' or 'gemini-3.1-pro'). Description says 'Model ID' but doesn't enumerate valid values or provide discovery. LLMs will hallucinate invalid model names. Should offer a list_models or validate against known registry.