CarbonCode exposes 3 tools with mixed quality. The 'add' tool is trivial but well-formed. 'semantic_search' has an excellent, detailed description (250+ chars) explaining WHEN to use it vs alternatives, with well-typed parameters and clear constraints. 'mcp_tools' is a meta-tool with good structure but lacks output schema documentation. Parameter descriptions are present and specific. However, no tool declares output schemas, limiting LLM ability to plan downstream calls. Error handling guidance is absent. The server is a CLI agent framework, not a traditional MCP server, which affects composition patterns.

• Document output schemas for all tools. For semantic_search, specify: {type: 'object', properties: {snippets: {type: 'array', items: {type: 'object', properties: {file: string, score: number, text: string}}}}}. For mcp_tools, document the structure returned by each action.
• Add error handling guidance to semantic_search and mcp_tools descriptions. E.g., 'If no results found, try lowering minScore or expanding topK. If the index is unavailable, fall back to search_content.'
• Expand 'add' description to explain context: 'Simple integer addition tool for calculator workflows. Use when the agent needs to perform arithmetic as part of a larger task.'
• Consider splitting mcp_tools into separate tools: mcp_list_tools, mcp_search_tools, mcp_describe_tool, mcp_call_tool. This reduces cognitive load and makes intent clearer to LLMs.
• Add per-parameter constraints to mcp_tools: document valid enum values for 'action', specify limit range (1-20), offset minimum (0), and clarify that 'arguments' is only used with action='call'.