Monitor Twitter Complaints About Your Brand with Python

Complaint Radar is a small Python script for support, product and marketing teams. Give it a brand name and it returns a short morning report of public Twitter complaints about it, sorted by theme and by how much attention each post got.

The problem

Customers often complain in public before they open a ticket. Reading every mention by hand takes time, and most mentions are not complaints at all. Teams need a quick daily view: what are people unhappy about, is it growing, and which posts are getting the most attention.

What the app does

• Takes a brand name, other names for it, and the brand's own support account to leave out.
• Searches recent public X posts that pair the brand with complaint words such as "not working", "refund" or "can't log in".
• Removes reposts, duplicate posts and the brand's own posts.
• Sorts each post into a theme: outage, account, billing, bug, support, feature request or other.
• Writes a Markdown report: posts per theme, volume per day, the top posts by engagement, and the loudest posts per theme.
• Saves a JSON snapshot with every post and every API call, plus the total cost of the run.

How it works

1. Build the search. The script turns your input into one X search query, for example (PlayStation OR PSN) ("not working" OR refund OR outage ...) -from:AskPlayStation. The query supports X search operators such as OR, quotes, -word and from:.
2. Search one day at a time. For each day it calls GET /twitter with sort=Latest and a start and end date. If a page is full, it asks for the next page.
3. Optional replies. For the top complaint posts it calls GET /twitter/replies/post to show what other people answered.
4. Sort locally. Simple keyword rules assign a theme. Engagement score is likes + 2 x replies + reposts + quotes + bookmarks.
5. Write the report and add up the cost from the X-Desearch-Cost-Usd response header.

The code

The core call is one HTTP request:

    def get(self, path, params, label):
        if self.spent() >= self.max_spend:
            print(f"spend cap ${self.max_spend} reached, skipping {label}", file=sys.stderr)
            return None
        for attempt in (1, 2):
            when = time.strftime("%Y-%m-%d %H:%M:%S %z")
            t0 = time.perf_counter()
            try:
                r = self.s.get(BASE + path, params=params, timeout=120)
                status, hdr, text = r.status_code, r.headers, r.text
            except Exception as e:  # network error
                status, hdr, text = None, {}, str(e)
            lat = round(time.perf_counter() - t0, 2)
            try:
                body = json.loads(text)
            except ValueError:
                body = {"_raw_text": text[:2000]}
            items = body if isinstance(body, list) else []
            c = {"when": when, "label": label, "path": path, "params": params, "status": status, "latency_s": lat,
                 "items": len(items), "units": int(hdr.get("X-Desearch-Usage-Count") or 0) or None,
                 "cost_usd": float(hdr.get("X-Desearch-Cost-Usd") or 0), "attempt": attempt}
            self.calls.append(c)
            print(f"[{when}] {status} {path} {json.dumps(params)} items={c['items']} units={c['units']} "
                  f"cost=${c['cost_usd']:.5f} {lat}s", file=sys.stderr)
            if self.raw_dir:
                s = json.dumps({"call": c, "request": {"url": BASE + path, "params": params, "headers": {"Authorization": "<redacted>"}},
                                "response": body}, ensure_ascii=False, indent=1)
                assert self.key not in s
                open(os.path.join(self.raw_dir, f"{len(self.calls):03d}-{label}.json"), "w").write(s)
            if status == 200:
                return body
            if status in (429, 500, 502, 503, 504, None) and attempt == 1:
                time.sleep(2)
                continue
            return None

The repo folder has one script, complaintradar.py (about 380 lines: query builder, day-by-day paging, themes, report writer), and offline tests in test_complaintradar.py.

Endpoints

• GET /twitter
• GET /twitter/replies/post

Example output

From a run for PlayStation on 2026-10-08, covering 3 days (2026-10-06 to 2026-10-08 UTC). 123 posts after removing duplicates.

The full report also lists posts per day, the 10 complaint posts with the most engagement, and the loudest posts in each theme, each with a link to the post.

Run it yourself

The public examples folder linked above includes the script and a README. Install its requirements, set DESEARCH_API_KEY from console.desearch.ai, and use the run command in that README.

Useful flags: --days (how many days back), --keywords (your own complaint words), --exclude (words to filter out), --max-spend (stop at a dollar limit, default $0.10).

What it costs

The example run above made 10 API calls (7 searches returning 127 posts and 3 reply lookups) and cost $0.028, read from the X-Desearch-Cost-Usd header on 2026-10-08. A full page of 20 posts cost $0.003. See pricing.

Ideas to extend it

• Run it every morning with cron and send the report to a team chat.
• Track two or three brands side by side to compare complaint themes.
• Replace the keyword themes with an LLM classifier for more nuance.
• Alert when a theme jumps above its usual daily count.

{
  "mcpServers": {
    "desearch": {
      "url": "https://mcp.desearch.ai/mcp",
      "headers": {
        "Authorization": "Bearer <DESEARCH_API_KEY>"
      }
    }
  }
}

Use Desearch to find public X posts from the last day that complain about the brand I name. Group them by theme and quote the posts you retrieve. Skip posts that are not complaints. Do not invent posts.