X Research

X Research (xAI/Grok)

AI-powered research using xAI's Grok API. Returns AI summaries with citations, not raw tweet data.

Want raw tweets instead? Use these skills:

• x-user-timeline - Raw tweets from a user
• x-tweet-search - Raw search results
• x-tweet-fetch - Raw single tweet
• x-user-lookup - User profile data

When tools from X's official XMCP server are available, prefer them for raw X data. XMCP normally runs locally at http://127.0.0.1:8000/mcp; https://docs.x.com/mcp is the separate hosted documentation server. Otherwise use the skills above. Both raw-data routes require X credentials; this synthesis skill uses XAI_API_KEY instead.

When to Use

USE THIS SKILL FOR:

• Summarized research with AI analysis
• Current events and breaking news
• Social sentiment ("What are developers saying about Y?")
• X/Twitter trending topics and viral posts
• Emerging tools/frameworks not in documentation yet
• Community opinions on best practices

DON'T USE FOR:

• Raw tweet data (use x-user-timeline, x-tweet-search)
• Specific user's posts verbatim (use x-user-timeline)
• Well-documented APIs (use WebFetch instead)

Setup Requirements

Check if the API key is set:

if [ -z "$XAI_API_KEY" ]; then
  echo "ERROR: XAI_API_KEY is not set"
  echo "1. Get API key from https://x.ai/api"
  echo "2. Add to profile: export XAI_API_KEY=\"your-key\""
  echo "3. Restart terminal and Claude Code"
  exit 1
else
  echo "XAI_API_KEY is configured"
fi

If unavailable, inform the user that XAI_API_KEY must be configured before using this skill.

API Overview

Endpoint: https://api.x.ai/v1/responses

Resolve the loaded skill directory once per shell call, then use its request wrapper. For Claude Code plugins, derive it from CLAUDE_PLUGIN_ROOT. For Codex, use the absolute directory of this loaded SKILL.md:

X_RESEARCH_SKILL_DIR="${CLAUDE_PLUGIN_ROOT}/skills/x-research"
"$X_RESEARCH_SKILL_DIR/scripts/research.sh" <<'JSON'
{"input":"[YOUR QUERY]","tools":[{"type":"web_search"},{"type":"x_search"}]}
JSON

Leave XAI_RESEARCH_MODEL unset, or set it to auto, latest, or grok-latest/grok-4-latest, to select the newest canonical general-purpose Grok model available to the current API key. The selector currently resolves that policy to grok-4.5; it advances automatically when xAI publishes a newer canonical model. Do not trust an alias merely because its name contains latest—the live catalog can retain a lagging alias. Set a versioned ID such as grok-4.5, or a versioned alias such as grok-4.5-latest, only to request a deliberate pin. The selector verifies explicit pins and always emits the canonical model ID. The wrapper selects and calls the model in one process, so model choice cannot disappear between separate shell tool calls. It accepts a request JSON object on stdin, overwrites any embedded model, and returns the raw Responses API JSON.

Available Search Tools:

• web_search - Searches the internet and browses web pages
• x_search - Semantic and keyword search across X posts

Basic Usage

Simple Query with Search Tools

"$X_RESEARCH_SKILL_DIR/scripts/research.sh" <<'JSON' \
  | tr -d '\000-\010\013\014\016-\037' \
  | jq -r '[.output[] | select(.type=="message")][-1].content[0].text'
{
    "input": [{"role": "user", "content": "[YOUR QUERY]"}],
    "tools": [{"type": "web_search"}, {"type": "x_search"}]
}
JSON

Query with Usage Tracking (Recommended)

RESPONSE=$("$X_RESEARCH_SKILL_DIR/scripts/research.sh" <<'JSON'
{
    "input": [{"role": "user", "content": "[YOUR QUERY]"}],
    "tools": [{"type": "web_search"}, {"type": "x_search"}]
}
JSON
)

# Extract usage stats
TOOL_CALLS=$(echo "$RESPONSE" | jq -r '.usage.num_server_side_tools_used // 0')
COST_TICKS=$(echo "$RESPONSE" | jq -r '.usage.cost_in_usd_ticks // 0')
COST_USD=$(jq -n --argjson ticks "$COST_TICKS" '$ticks / 10000000000')
echo "Tool calls: $TOOL_CALLS | Estimated cost: \$$COST_USD"
echo ""
echo "$RESPONSE" | tr -d '\000-\010\013\014\016-\037' | jq -r '[.output[] | select(.type=="message")][-1].content[0].text'

Search Tool Parameters

Web Search Options

"tools": [{
  "type": "web_search",
  "allowed_domains": ["github.com", "stackoverflow.com"],
  "excluded_domains": ["pinterest.com"],
  "enable_image_understanding": true
}]

• allowed_domains - Restrict to these domains only (max 5)
• excluded_domains - Exclude these domains (max 5)
• enable_image_understanding - Analyze images found during search

X Search Options

"tools": [{
  "type": "x_search",
  "allowed_x_handles": ["elikimonimus", "anthropaboris"],
  "excluded_x_handles": ["spambot"],
  "from_date": "2025-01-01",
  "to_date": "2025-01-16",
  "enable_image_understanding": true,
  "enable_video_understanding": true
}]

• allowed_x_handles - Only include posts from these accounts (max 20)
• excluded_x_handles - Exclude posts from these accounts (max 20)
• from_date / to_date - ISO8601 date range (YYYY-MM-DD)
• enable_image_understanding - Analyze images in posts
• enable_video_understanding - Analyze videos in posts

Examples

X/Twitter Trending Topics

"$X_RESEARCH_SKILL_DIR/scripts/research.sh" <<'JSON' \
  | tr -d '\000-\010\013\014\016-\037' \
  | jq -r '[.output[] | select(.type=="message")][-1].content[0].text'
{
    "input": [{"role": "user", "content": "What is currently trending on X? Include viral posts and major discussions."}],
    "tools": [{"type": "x_search"}]
}
JSON

Developer Sentiment on a Topic

"$X_RESEARCH_SKILL_DIR/scripts/research.sh" <<'JSON' \
  | tr -d '\000-\010\013\014\016-\037' \
  | jq -r '[.output[] | select(.type=="message")][-1].content[0].text'
{
    "input": [{"role": "user", "content": "What are developers saying about [TOPIC] on X? Include recent discussions and opinions."}],
    "tools": [{"type": "x_search"}, {"type": "web_search"}]
}
JSON

News and Web Research

"$X_RESEARCH_SKILL_DIR/scripts/research.sh" <<'JSON' \
  | tr -d '\000-\010\013\014\016-\037' \
  | jq -r '[.output[] | select(.type=="message")][-1].content[0].text'
{
    "input": [{"role": "user", "content": "Latest news and developments about [TOPIC]"}],
    "tools": [{"type": "web_search"}]
}
JSON

Historical Research with Date Range

"$X_RESEARCH_SKILL_DIR/scripts/research.sh" <<'JSON' \
  | tr -d '\000-\010\013\014\016-\037' \
  | jq -r '[.output[] | select(.type=="message")][-1].content[0].text'
{
    "input": [{"role": "user", "content": "What happened with [TOPIC] from January 9 through January 16, 2025?"}],
    "tools": [{
      "type": "x_search",
      "from_date": "2025-01-09",
      "to_date": "2025-01-16"
    }, {
      "type": "web_search"
    }]
}
JSON

Response Structure

The response contains:

{
  "output": [
    {"type": "web_search_call", "status": "completed", ...},
    {"type": "custom_tool_call", "name": "x_semantic_search", ...},
    {
      "type": "message",
      "content": [{
        "type": "output_text",
        "text": "The actual response with [[1]](url) inline citations",
        "annotations": [{"type": "url_citation", "url": "...", ...}]
      }]
    }
  ],
  "citations": ["url1", "url2", ...],
  "usage": {
    "input_tokens": 11408,
    "output_tokens": 2846,
    "num_server_side_tools_used": 8,
    "cost_in_usd_ticks": 429052500
  }
}

Extracting content: jq -r '[.output[] | select(.type=="message")][-1].content[0].text' — select the last MESSAGE element explicitly; on grok-4.5 the final output[] element is a reasoning block, and reasoning summaries can contain raw control characters that break strict JSON parsing of the whole payload. The tr strip in every example removes raw control characters grok-4.5 reasoning summaries can emit, which otherwise break strict JSON parsing of the whole payload before jq ever selects.

Extracting citations: jq -r '.citations[]'

Pricing

Search and token pricing changes independently of this skill. Read the exact charge from response.usage.cost_in_usd_ticks; xAI defines one USD as 10,000,000,000 ticks. Consult the provider pricing page before estimating future requests.

Error Handling

If the API returns an error:

1. Check XAI_API_KEY is set and valid
2. Verify the model resolves for the current API key with GET /v1/models
3. Check rate limits haven't been exceeded
4. Ensure tools array is properly formatted
5. Report the error to the user with the API response

Common errors:

• "Internal error" - Usually wrong model name or malformed request
• Timeout - Request took too long (rare with agentic API)

Migration Notes

DEPRECATED (January 12, 2026): The old Live Search API using search_parameters is being deprecated. This skill uses the new agentic tool calling API which is faster and more capable.

Old format (deprecated):

{
  "model": "grok-4-latest",
  "messages": [...],
  "search_parameters": {"mode": "on"}
}

New format (use this after verifying a current model ID):

{
  "model": "<verified-current-model-id>",
  "input": [...],
  "tools": [{"type": "web_search"}, {"type": "x_search"}]
}

References

• LLMs.txt (AI-optimized docs): https://docs.x.com/llms.txt
• xAI API Documentation
• Search Tools Guide
• Models Reference
• X.com API v2