dragonkid/grok-query

Grok Query via agent-browser

Query X's Grok AI through browser automation to get analysis backed by real-time X post data.

Grok has direct access to X/Twitter's full post corpus — trending discussions, sentiment shifts, community reactions, breaking narratives. For crypto sentiment, breaking news reactions, community FUD/FOMO analysis, or any topic where X discussions are the primary signal, Grok provides data that web search tools cannot reach.

Prerequisites

• User's Chrome must have remote debugging enabled (chrome://inspect/#remote-debugging checkbox) and be logged into X
• agent-browser CLI must be installed

Workflow

Step 1: Establish CDP connection

Chrome 144+ shows an "Allow remote debugging?" dialog on every new CDP WebSocket connection. Never use --auto-connect open <url> — it bundles connection + navigation, and if the user is slow to click Allow, agent-browser times out and retries, stacking dialogs that make buttons appear unresponsive.

Always connect separately first, then navigate:

WS_PATH=$(sed -n '2p' ~/Library/Application\ Support/Google/Chrome/DevToolsActivePort)
agent-browser connect "ws://127.0.0.1:9222${WS_PATH}"

The user clicks Allow on Chrome's dialog. Once connect succeeds, all subsequent commands in this session work without further dialogs.

Step 2: Open Grok in a new tab

Always open a new tab for each query — this preserves the user's existing tabs and allows clean cleanup after extraction.

agent-browser tab new "https://x.com/i/grok"
agent-browser wait --load networkidle

To resume a previous conversation:

agent-browser tab new "https://x.com/i/grok?conversation=<ID>"
agent-browser wait --load networkidle

Step 3: Submit the question

Snapshot to find the input field, fill it, then re-snapshot to find the submit button.

Refs shift after fill — the ref you found before filling will point to a different element after. This is the most common source of bugs: clicking "专家" instead of "问 Grok 问题" because you used a stale ref.

# Find input, fill question
agent-browser snapshot -i -c | grep -E "textbox"
agent-browser fill @<input-ref> "<question>"
# MUST re-snapshot — refs have changed
agent-browser snapshot -i -c | grep -E "button.*(问|Grok 问题|submit)"
# Scroll into view and click — sometimes the button is obscured
agent-browser scrollintoview @<submit-ref>
agent-browser click @<submit-ref>

Do not press Enter — Grok's input intercepts it for newlines.

Step 4: Wait for completion, then extract

Don't use keyword detection ("思考"/"Thinking") — these words can appear in Grok's response text or cited tweets and cause the poll to never exit.

Instead, use content stabilization: poll until two consecutive reads return the same content, meaning Grok has finished writing.

sleep 10
PREV=""
for i in $(seq 1 24); do
  CURR=$(agent-browser eval --stdin 2>&1 <<'JSEOF'
(() => (document.querySelector('main')?.innerText || '').length.toString())()
JSEOF
  )
  if [ -n "$PREV" ] && [ "$CURR" = "$PREV" ]; then
    echo "Content stabilized at poll $i (length: $CURR)"
    break
  fi
  PREV="$CURR"
  sleep 5
done

Then extract the full response:

agent-browser eval --stdin <<'JSEOF'
(() => {
  const main = document.querySelector('main');
  return main ? main.innerText : 'ERROR: main element not found';
})()
JSEOF

If the response seems truncated (ends mid-sentence), scroll down and re-extract:

agent-browser scroll down 5000
sleep 2
# re-run the eval above

Step 5: Record conversation ID and close the tab

agent-browser eval --stdin <<'JSEOF'
(() => window.location.href)()
JSEOF

Save the conversation ID from x.com/i/grok?conversation=<ID> — use it for follow-ups instead of starting new conversations.

Close only the Grok tab (not the entire browser session):

agent-browser tab close

This leaves the CDP connection alive and the user's other tabs untouched. Use agent-browser close only when you're done with all browser automation for the session.

Formatting the Response

The raw main.innerText includes sidebar navigation mixed in. Clean it up:

1. The Grok response starts after the user's question text (which appears first)
2. It ends before the source counts (e.g., "8 网页") and suggested follow-up prompts
3. Strip lines that are clearly navigation: "主页", "探索", "通知", "聊天", "Grok", "书签", "创作者工作室", "Premium", "个人资料", "更多", "发帖", "DragonKid", "@username", "查看新帖子", "专家"
4. Preserve structure (headings, bullets, bold, tables)
5. Note cited sources (X post count + web page count shown at end)
6. Add a note that this is Grok's analysis based on X data as of the current date

Follow-up Questions

To continue in the same conversation:

1. Establish connection (Step 1)
2. agent-browser tab new "https://x.com/i/grok?conversation=<ID>"
3. Same fill → re-snapshot → click submit flow
4. Extract and format response the same way
5. agent-browser tab close when done

Troubleshooting

| Problem | Solution | |---------|----------| | Dialog buttons unresponsive | Multiple dialogs stacked — agent-browser close, then reconnect with Step 1 | | Connection timeout | Verify chrome://inspect/#remote-debugging checkbox is enabled | | DevToolsActivePort not found | Chrome not running or remote debugging not enabled | | Clicked wrong button | Refs shifted after fill — always re-snapshot before clicking | | Poll loop never exits | Content length stabilization should handle this; if not, increase sleep interval | | Partial/truncated response | Scroll down 5000px then re-extract | | Empty main.innerText | Page still loading — agent-browser wait --load networkidle then retry | | Lost conversation | Use saved conversation ID, or click "聊天历史记录" button |