MCP Gateway exhibits mixed quality across its 9 tools. Tool naming follows verb_noun conventions (searchTools, getToolSchema, executeCode, callTool, createSkill, listSkills, executeSkill, deleteSkill), which is strong. However, descriptions vary significantly in quality and completeness. Most tools have basic descriptions (50-150 chars), but several lack crucial context about state modification, error handling, and when to use each tool. Input schemas are present and well-typed with enums and constraints (e.g., detailLevel in searchTools, mode in getToolSchema), but output schemas are undocumented, a critical gap for agent planning. Parameter descriptions are present but inconsistent in depth. The executeCode tool (marked IRREVERSIBLE) lacks explicit confirmation/dry-run patterns despite its destructive capability. Error handling guidance is absent across all tools. Security concerns: executeCode has no visible sandboxing validation in the schema definition itself, though source hints at VM isolation.

• Document output schemas for all 9 tools. For searchTools, specify: {tools: [{name, description, backend, category, estimatedTokens}], total: number, hasMore: boolean}. For executeCode, specify: {result: string, executionTimeMs: number, error?: {message, type}}.
• Add confirmation/dry-run modes to executeCode and executeSkill. Accept a 'dryRun' boolean parameter (default false). When true, parse and validate the code/skill without executing. This prevents accidental side effects.
• Expand error handling in tool descriptions. For executeCode, add: 'Returns error details if code times out (>timeoutMs), contains forbidden globals, or throws an uncaught exception. Suggest reducing code complexity or increasing timeout if retrying.'
• Enhance parameter descriptions for callTool and executeSkill. Instead of 'Tool arguments', specify: 'Object mapping parameter names to values. Must match the schema of the target tool, invalid params cause execution to fail. Call getToolSchema first if unsure of required params.'
• Add an 'estimatedTokens' field to getToolSchema output to help agents budget context. Compute using ToolDiscovery.estimateTokens() and include in the response.
• Document how searchTools output (tool names) chains to callTool input (toolName parameter). Include an example: 'searchTools with query='send' returns [{name: 'sendEmail', ...}]. Pass name to callTool({toolName: 'sendEmail', arguments: {...}})'.
• Implement and document rate limiting for executeCode and callTool to prevent runaway agents from overwhelming the server or downstream backends. Return a 429 error with retry guidance when limits are exceeded.

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

• Relies on Stateful initialize / Mcp-Session-Id (removed; protocol is stateless) - make each request self-contained