xAI Grok Search

**Judgement:** Solid, focused integration for real-time web and X search via xAI Grok. Very useful when you explicitly need fresh information, social sentiment, or constrained-domain search, but it relies heavily on the upstream API’s behavior and lacks robust error/safety layers. **Key strengths:** - Direct use of xAI’s official `web_search` and `x_search` tools through `/v1/responses` with a clean, minimal wrapper. - Good control over scope (allowed/excluded domains and X handles, date ranges) to improve relevance. - Optional image and video understanding for richer analysis. - Clear documentation, examples, and operational tips (rate limiting, key setup, query scoping). **Main risks / limitations:** - **Error handling & robustness:** No explicit checks for HTTP errors, malformed responses, or missing fields; assumes `data.output` and `data.citations` always exist and be well-formed, which can cause runtime failures if the API changes or misbehaves. - **Safety & content quality:** No moderation, filtering, or prompt-injection defenses. Results may surface harmful, biased, or unverified content directly from the web/X; the caller must handle safety and trust. - **Latency & cost:** Default model (`grok-4-1-fast-reasoning`) can be slow (30–60+ seconds) and more expensive; there is no built-in model auto-selection or caching beyond what the docs recommend externally. - **Coupling to xAI response shape:** Strong assumptions about the structure of `data.output` (last element is the answer) and `data.citations`, which could break if the API evolves. **Recommended scenarios:** - When your agent needs **fresh news, live data, or social chatter** that a static model cannot reliably provide. - When you want **citation-backed answers** and can tolerate variability and occasional failures from an external API. - When you need **targeted web/X research** (e.g., only from specific domains or accounts) and are prepared to implement your own safety, retry, and caching layers on top of this skill. Less suitable for: - Purely historical or general-knowledge questions where the base model is sufficient. - Safety-critical or compliance-sensitive use cases without additional moderation and verification. - Latency-sensitive workflows that cannot tolerate slower reasoning searches or transient API slowdowns.