This MCP server exposes 16 tools with significant quality gaps across naming, descriptions, and schemas. Most tools have basic descriptions (30-80 chars) but lack detailed parameter documentation, proper error handling guidance, and comprehensive output schema definitions. Tool names follow verb_noun conventions but many are compound operations (codex-reply, codex-start, codex-status) that could be clearer. Parameter descriptions exist but are often minimal (5-20 chars: 'ID of the job', 'Offset in the result'). Critical issue: No visible input schema validation or output schema documentation in the source code, schemas are inferred from the provided tool spec but not verifiable in server.js implementation. Security concern: The server wraps external CLI tools (Codex, Gemini, Claude) and passes prompts directly without visible sanitization or injection prevention.

• Add output schema documentation for every tool. Document response field names, types, and meanings. E.g., codex_status should return {jobId, status, progress, completionPercentage, error?}.
• Expand parameter descriptions to 50 - 100 characters with actionable detail. For offset/cursor parameters, document pagination semantics: 'Offset (0-based position in result stream; omit for first page)'. For enum parameters like sandbox, explain security implications: 'Sandbox mode: read-only blocks file writes; workspace-write allows writes to designated dir; danger-full-access grants full system access (use with caution)'.
• Rename bare noun tools. Change 'gemini' to 'run_gemini' and 'codex' to 'run_codex_interactive'. Use underscore separators instead of hyphens (codex-reply → codex_reply, remote-browser → remote_browser).
• Add tool descriptions that explain WHEN to use each tool. E.g., 'codex: Synchronous execution of a prompt in an interactive Codex workspace (blocks until completion). Use for quick operations under 5 minutes. For longer tasks, use codex_start (async).' Similarly, distinguish between codex_reply (continue existing conversation, synchronous) and codex_reply_start (async variant).
• Document error handling. For each tool, add a sentence like: 'Returns error code TIMEOUT if the job exceeds [time limit]. Returns INVALID_SANDBOX if the specified sandbox mode is not available. See [error reference] for recovery steps.' Include actionable error messages in implementation.
• Add tool annotations (readOnlyHint, destructiveHint, idempotentHint) to tool definitions. Mark codex_cancel and remote_browser as destructiveHint=true. Mark codex_status and codex_result as readOnlyHint=true.