Loading ToolBench
Evaluates multi-channel ROAS performance and shifts budget to the highest return channels
This MCP server has severe definition quality issues that make it unsuitable for production use. The single tool 'execute_skill_action' lacks a properly documented input schema, has inadequate parameter descriptions, and is missing critical error handling guidance. The implementation appears to be a skeleton MCP wrapper around a working Python optimization algorithm, but the tool interface does not meet minimum standards for LLM-driven composition.
Executes ad spend optimization based on multi-channel ROAS performance
Input schema for 'channel_performance' parameter is empty, declares type 'object' but provides NO 'properties' field, making it impossible for LLMs to understand what structure is required. According to the rubric, missing or incomplete schemas for parameters result in a schema score of 0.
Parameter 'channel_performance' description ('Dictionary of channel performance data with ROAS values') is vague and provides no guidance on required fields, expected keys, or value constraints. Example from code shows 'google', 'meta', 'tiktok' as keys with nested 'spend' and 'roas' fields, but this structure is not documented in the schema or description.
No output schema documented. The tool returns a dict with 'allocated_budget' (dict of channel→float), 'projected_revenue' (float), and 'blended_roas' (float), but this structure is not visible in the tool definition. LLMs cannot plan downstream tool calls or extract the right data without a documented return type.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 29 | 2026-07-28+ | v2 |
Tool name 'execute_skill_action' is generic and does not follow verb_noun conventions clearly (e.g. 'optimize_budget' or 'allocate_ad_spend' would be more descriptive). The current name does not convey that this is specifically about ad spend allocation and ROAS optimization.
No error handling guidance. The tool description does not explain what happens if channel_performance is empty, if ROAS values are missing, or if total_budget_usd is 0 or negative. The code shows a fallback (total_roas = 1.0 if sum is 0), but the LLM is unaware of this recovery behavior.
No parameter type enforcement or validation constraints visible. The schema shows 'total_budget_usd' as type 'number' but provides no minimum, maximum, or range guidance. LLMs may pass negative budgets, zero, or extremely large values without knowing the valid range.
The mcp_server.py implementation does not actually register the tool with an MCP-compliant tool definition. It only prints a static JSON status message. There is no evidence of a proper MCP server implementation that handles tool invocations, validates inputs, or returns structured responses according to the MCP protocol.