Loading ToolBench
A complete A2A Protocol toolset for seamlessly generating Servers and Clients with multi-agent orchestration, routing, and specialized agent implementations.
Critical failures across all dimensions. Tool names lack action verbs (analytics_handler, chart_handler, mcp_handler are generic handlers, not verbs). Input schemas are present but minimal, only 'parts' (array) and 'context' (object) with no type details or field descriptions. Descriptions are present but generic and under 100 chars, offering no guidance on WHEN to use each tool or what distinguishes them. No output schemas documented. No error handling guidance. No parameter constraints (enums, ranges). The 'parts' and 'context' parameters are undescribed, LLMs cannot infer what fields they contain or how to construct them. Tool composition is poor: handlers are tightly coupled to specific implementations (Ollama, Confluence, MCP) rather than abstracting the interface. No idempotency guarantees, no pagination, no recovery guidance.
Returns detailed health status of all registered agents
Processes data and generates analytics reports using LLM-based analysis
Generates visual charts via AntV MCP server, producing SVG and PNG visualizations
Creates and updates Confluence documentation pages via MCP interface
Fetches full agent card for a given alias
Returns a list of all registered agents
Handles long-running asynchronous tasks
Tool names are generic handlers (analytics_handler, chart_handler, mcp_handler) instead of action verbs. LLMs cannot infer intent from names alone. Should be: generate_analytics, create_chart, execute_mcp_tool, create_plan, generate_report, validate_content, analyze_multimodal, stream_results, execute_task, process_text.
Input schemas are minimal and undescribed. All handler tools accept only 'parts' (array) and 'context' (object) with no field-level documentation. LLMs cannot construct valid requests without knowing what fields 'parts' and 'context' contain, their types, or constraints. Schema score capped at 20 due to missing type details.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 27 | 2026-07-28+ | v2 |
Executes MCP tools using LLM-based tool calling and summarizes findings
Handles multimodal input (text and images) for vision-based analysis
Creates execution plans and DAGs for multi-agent orchestration based on user queries
Registers a new agent via their Agent Card with signature verification
Generates structured security audit reports from files or text input
Returns real-time streaming updates for long-running operations
Processes text-only queries and returns text responses
Reviews and validates content for accuracy, safety, and adherence to instructions
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields. Handlers return unstructured text responses (e.g., 'Analytics Report Generate by gemma3:270m: ...'). Should return structured JSON with typed fields (status, data, metadata, errors).
No error handling guidance. Handlers catch exceptions and return bare error strings (e.g., 'Analytics Error: ...'). No indication of whether errors are retryable, user-fixable, or fatal. No recovery suggestions. LLMs cannot self-correct.
Tool composition is poor. Handlers are tightly coupled to specific implementations (Ollama for analytics, AntV MCP for charts, Confluence API). No abstraction layer. If Ollama is unavailable, analytics_handler fails with no fallback. Breaks the single-responsibility principle.