The 5-minute HN monitoring setup that replaces a $5K SDR tool

Five minutes. One Algolia HTTP call. A cron job. An optional 30-line LLM classifier. That's the entire HN monitoring stack you need to replace the SDR intelligence tool your VP of Sales is about to expense for $5K a year. We've built this e

Five minutes. One Algolia HTTP call. A cron job. An optional 30-line LLM classifier. That's the entire HN monitoring stack you need to replace the SDR intelligence tool your VP of Sales is about to expense for $5K a year.

We've built this exact setup for hundreds of HN keyword monitors inside Shadow Inbox. It's not complicated. The complicated part is knowing which keywords to monitor and what to do with the results, both of which we'll cover.

HackerNews has a free, undocumented-but-stable search API. Most of the SDR tool category is reselling it back to you with a logo on top.

The whole stack, in five steps.

1. Algolia HN search API  (free, no auth)
       │
       ▼
2. Keyword + filter spec  (JSON, version-controlled)
       │
       ▼
3. Cron job or webhook    (every 15 min)
       │
       ▼
4. LLM intent filter      (optional, $0.002/post)
       │
       ▼
5. Slack / email destination  (Resend, webhook, whatever)

The first three steps are 5 minutes. Steps 4 and 5 add another 20 minutes if you've never wired up an LLM call before. The whole thing fits in a single Vercel cron + edge function, or a node cron.js on a $5 VPS.

Step 1: The Algolia HN search API.

HN runs on Algolia for search. The API is public, undocumented in any official place, and stable since 2014. The base endpoint:

https://hn.algolia.com/api/v1/search_by_date

Two query types matter:

• search — relevance ranked. Use for one-off queries.
• search_by_date — chronological. Use for monitoring (you want new stuff, not popular stuff).

The full param list is on the Algolia HN search docs page. The ones you'll actually use:

• query — your keyword string
• tags — filter by post type: story, comment, ask_hn, show_hn, front_page
• numericFilters — filter by timestamp, points, comment count
• hitsPerPage — default 20, max 1000

Step 2: The keyword + filter spec.

Don't hardcode keywords in your cron. Put them in a JSON file you can edit without redeploying.

{
  "monitors": [
    {
      "name": "self-hosted-analytics",
      "queries": ["plausible alternative", "self-hosted analytics", "umami vs"],
      "tags": "(story,ask_hn,comment)",
      "min_points": 1,
      "max_age_minutes": 30
    },
    {
      "name": "ci-pain",
      "queries": ["github actions slow", "circleci alternative"],
      "tags": "(comment)",
      "min_points": 0,
      "max_age_minutes": 60
    }
  ]
}

Two non-obvious choices baked into this shape:

Comments are first-class signals. Most monitoring tools only watch stories. The buyer who's actually frustrated is in the 47-comment Show HN thread saying "yeah we tried CircleCI and it was painful." Watch comments.

Max age, not min age. You want fresh posts, not popular ones. A post that was filtered out last poll cycle should not show up again 6 hours later when it hits the front page. Filter on created_at_i > now - 30min.

Step 3: The cron.

A 30-line Node script that hits the API, dedupes against a flat-file or Postgres seen_ids table, and pushes new hits to step 4.

// hn-monitor.ts — runs every 15 min
import { readFileSync } from "fs";
const config = JSON.parse(readFileSync("./monitors.json", "utf8"));
const seen = await loadSeenIds();
const now = Math.floor(Date.now() / 1000);
 
for (const monitor of config.monitors) {
  for (const q of monitor.queries) {
    const url = new URL("https://hn.algolia.com/api/v1/search_by_date");
    url.searchParams.set("query", q);
    url.searchParams.set("tags", monitor.tags);
    url.searchParams.set(
      "numericFilters",
      `created_at_i>${now - monitor.max_age_minutes * 60}`
    );
    const hits = (await fetch(url).then(r => r.json())).hits;
    for (const hit of hits) {
      if (seen.has(hit.objectID)) continue;
      await processHit(hit, monitor);
      seen.add(hit.objectID);
    }
  }
}

Stick that in a vercel.json cron, a GitHub Actions schedule, or a literal crontab -e entry. Whatever you have. Don't let infrastructure be the reason this doesn't ship.

Step 4: The LLM intent filter.

The Algolia API returns posts that match your keywords. Most of those posts will not be buyers. They'll be devs evaluating, students researching, hobbyists nerd-sniping. You need an LLM filter or you'll send yourself junk.

The shape of the prompt:

You are filtering HackerNews posts for buying intent.
The user sells: {one-line product description}.
 
Read the post below. Answer in JSON:
{
  "intent": "buyer_now" | "buyer_soon" | "evaluating" | "off_topic",
  "confidence": 0.0-1.0,
  "evidence": "quoted span from the post that justifies the verdict"
}
 
Only mark buyer_now if the OP states they are deciding within 30 days
and have a current pain. Otherwise default to evaluating or off_topic.

Run this with Claude Sonnet 4.6 at temperature 0. Cost is roughly $0.002 per post for typical HN body length. For 50 posts a day, that's $3 a month.

The chain-of-thought trick (forcing evidence quotation before classification) cuts false positives roughly in half. We cover the full LLM intent classifier design in our buying intent piece.

Step 5: Push to where humans look.

Don't build a dashboard you'll never check. Push to where you already work.

• Slack — incoming webhook, one channel per monitor. Format: title link, score, evidence quote, "reply" button.
• Email — Resend or your SMTP. Daily digest at 8:30am with everything from the last 24 hours sorted by intent score.
• Inbox — your existing CRM via Zapier or a direct API call.

The right destination depends on the decay curve. HN posts decay slower than Reddit (median useful-comment window is 18 hours, not 6), so a daily email digest works. If you're chasing breaking Show HN launches, use Slack.

The full multi-channel sequencing thinking — what to do once a signal lands somewhere — is in the multi-channel sequencing playbook.

A working example: monitoring CI/CD pain.

Suppose you sell a faster CI runner. Here's the actual config we'd run:

{
  "monitors": [
    {
      "name": "ci-frustration",
      "queries": [
        "github actions slow",
        "ci pipeline slow",
        "circleci alternative",
        "ci minutes expensive",
        "buildkite vs"
      ],
      "tags": "(story,comment,ask_hn)",
      "max_age_minutes": 30
    }
  ]
}

Five keywords, hits posts and comments, polled every 15 minutes. In a typical week this surfaces 8 to 15 raw matches, of which the LLM filter promotes 2 to 4 to buyer_now. That's a small enough volume that a founder can write personal replies without burning out.

The HN-specific intent playbook — Show HN, Ask HN, Who's Hiring, and the gold most people miss — is laid out in the HN intent piece. For comment-mining specifically, the HN comments piece covers the etiquette of replying without getting flamed.

What this stops working when.

Three failure modes worth naming:

Your category has no shared vocabulary on HN. If your buyers don't use the words you'd search for, the keyword monitor finds nothing. Sell to non-technical buyers? HN isn't the channel.

You set keywords too broad. "API" matches 80 posts a day. "REST API rate limiting frustration" matches one post a month. Start narrow and broaden.

You don't act within 24 hours. HN threads die. If your monitor fires at 6pm Friday and you check Monday morning, the OP has moved on. At minimum, push the signal somewhere you'll see it on weekends if you care about weekend posts.

Where Shadow Inbox fits.

The setup above is the version you build yourself in an afternoon. Shadow Inbox is the version with the intent filter pre-built, the Slack/email destinations wired up, the dedupe and decay timer baked in, and the same pipeline running on Reddit (and soon LinkedIn, X, Quora). If you'd rather skip the wiring, that's what we do. If you'd rather own the stack, the JSON above is what we'd build again.