The Agent Sandbox MCP server exposes only 2 tools with minimal documentation and no visible input schemas in the provided source. Tool names lack action verbs and are ambiguous. Descriptions are present but generic and lack actionable detail. No input parameter descriptions visible. No output schemas documented. The server appears to be a proxy/gateway pattern rather than implementing concrete tools, which limits its ability to follow standard tool design patterns. Code review of `pkg/handler/sbtool_handler.go` reveals tool registration but the actual schema definitions are not visible in the provided source material.

• Rename tools to action-verb format: 'sandbox-executor' → 'execute_sandbox_tool', 'sandbox-tools' → 'list_sandbox_tools'. This immediately clarifies intent for LLM tool selection.
• Add comprehensive parameter descriptions for each input. For 'sandbox_name': 'The unique identifier or name of the sandbox environment (e.g., "prod-analysis-1"). Available sandbox names can be retrieved via list_sandbox_tools().' For 'tool_name': 'The name of the tool to invoke within the sandbox (e.g., "bash-executor", "python-repl"). Must match a tool returned by list_sandbox_tools() for this sandbox.' For 'arguments': 'A JSON object containing tool-specific arguments. Structure depends on tool_name, refer to list_sandbox_tools() output for schema of each tool.'
• Document output schemas explicitly. For 'sandbox-executor': 'Returns {"success": boolean, "output": string|object, "error": string|null, "execution_time_ms": number}'. For 'sandbox-tools': 'Returns {"tools": [{"name": string, "description": string, "input_schema": object}], "total": number}'.
• Add error handling section to descriptions. E.g., 'Errors: Returns 404 if sandbox_name does not exist (try list_sandboxes() to enumerate available). Returns 400 if tool_name is not found in this sandbox (check list_sandbox_tools output). Returns 500 if execution times out (increase timeout parameter if sandbox tool supports it).'
• Consider decomposing the proxy pattern. Instead of a meta-tool that accepts arbitrary tool names, publish each sandbox tool individually as a separate MCP tool. This allows LLMs to see concrete schemas and make informed decisions about which tool to invoke.

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry