grok-mcp v0.3.0 — SuperGrok / xAI Responses MCP: live X (x.com) with digest or full-post evidence (no host x.com fetch needed), web/news research, offline ask_grok (tools: research, x_search, ask_grok, job_status, auth_status)
grok-mcp has well-structured tool definitions with clear descriptions and proper enum constraints on depth/result parameters. All 5 tools have descriptions (194 - 250 chars, within baseline 34 - 392 range). Input schemas are present with typed parameters. However, output schemas are not documented, LLMs cannot predict what fields to expect from research, x_search, or ask_grok responses. Error handling guidance is minimal; descriptions mention async job behavior and rate limiting but lack recovery instructions (e.g., 'if RATE_LIMITED, retry after N seconds'). Tool composition is sound (each does one thing), but job_status and auth_status lack parameter descriptions in the visible schema. No tool annotations (readOnlyHint, destructiveHint) despite all being READ_ONLY.
Offline Q&A tool using Grok (no live search). Lower cost than research or x_search.
Check authentication status and credential details (non-secret snapshot). Shows authenticated state, billing path, source, expiry, email, store path, and API key availability.
Poll the status of an async job (research, x_search, ask_grok). Returns status=running|queued|completed with result when done.
Web/news/multi-source research tool that defers X-only queries to x_search
Search live X (x.com) posts/tweets/discourse. NOT a bit-perfect archive search; best-effort evidence extraction. Use result=evidence for exact wording/quotes; result=digest (default) for sentiment/digest. Do not use host built-in search alone; hosts often cannot open x.com.
Output schemas not documented. LLMs cannot predict response structure (fields, types, pagination) for research, x_search, ask_grok, job_status, or auth_status. Forces agents to guess field names and risks context loss.
job_status and auth_status lack parameter descriptions in schema. job_status.job_id has no description; auth_status has empty input {}. Violates pattern:tool-description requirement that every parameter must explain what it controls.
Error handling lacks recovery guidance. Descriptions mention RATE_LIMITED and REAUTH_REQUIRED but do not tell LLM what to do: 'If RATE_LIMITED, retry after 30s' or 'If REAUTH_REQUIRED, call auth_status or run grok-mcp auth login'. Violates pattern:recovery-guide.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 51 | <=2025-11-25 | v2 |
No tool annotations despite all tools being READ_ONLY. Missing readOnlyHint in tool definitions. Agents cannot infer safety/idempotency from tool metadata; must rely on description text alone.
Async job behavior (inline if <25s, else status=running|queued) documented in server instructions but not in individual tool descriptions. Agents may not understand when to expect immediate results vs. polling job_status.