Sales Watchtower

A free, self-hosted account-intelligence dashboard: type in the names of companies you want to track, and it watches the news for you. Every article about a company on your list gets read by an AI, scored for how likely it is to be worth a sales outreach, and shown to you in a clean feed — no CRM, no manual searching, no spreadsheets.

Built with React, Vite, and Tailwind CSS v4. Works for any industry, any region, any product you sell — there's nothing hardcoded to one company, one vertical, or one country.

What it actually does

1. You add a company — type a name into "Track a Company" (on the My Accounts page). That's the entire setup. Optionally mark it as a customer, a prospect, or a competitor, and jot down a note on why you're tracking it.
2. It watches the news for that company, on a schedule (daily by default, hourly if you self-host).
3. An AI reads every new article about it, summarizes it, and scores it 1–5 on how strong an outreach trigger it is — a new executive hire, a funding round, an expansion into a new market, a public complaint, a regulatory issue, and so on.
4. You see it in the dashboard — a feed, grouped by company, with a plain-English "why this matters" explanation and a suggested angle for reaching out.

Nothing is invented about a real company: every synthesized line traces back to a real, cited news article. Contacts, tech stack, and financial data are deliberately not guessed at — those panels say what real data source would be needed to fill them in, rather than making something up.

Structure

A top nav of destinations, with signals opening in a slide-over drawer from anywhere.

Radar — the landing screen: what the pipeline found this week, the stages behind it (explained honestly — what's live vs. not yet built), and your top-priority accounts.

Global Feed — signals nested under the account they name, rather than a flat chronological list. Filter by type, priority, and date range; tab across signal categories; export the filtered set as CSV.

Competitors — companies you've flagged as competitors (not prospects) get their own page, read for a different purpose: not "how do I sell into this," but "what does this mean for how I position against them."

Accounts — the full derived account table: every company that's shown up in the news, whether you're tracking it yet or not. Sortable, with a 0–100 score and a P1–P3 priority tier.

My Accounts — your tracked companies, plus the "Track a Company" flow that's the whole point of this app: a name and (optionally) a note on why you're tracking it. A company you add here starts being watched by the next scheduled run — no signal has to exist yet.

Account detail — signal timeline, a "what you need to know" summary, the full signal list, a value pyramid, discovery-question prompts, and the score breakdown shown in the open (not hidden), plus every source.

Search — full-text across headlines, summaries, entities and company names, with saved alerts that re-run live against the current signal set.

Throughout: a Cmd/Ctrl+K command palette, dark/light theme, and a "synced Xh ago" indicator that turns amber if a scheduled run goes quiet or fails.

Account scoring

Accounts are derived from signals, not stored separately: every signal already carries matchedCompanies and entity, which is everything a rollup needs. A signal naming a company on your tracked list makes that account "tracked"; a signal naming an unrecognized company still produces a row — that's the whitespace view, a company worth adding that you haven't yet.

The score weights the best outreach trigger (55%), how warm the account is right now (30%), and sustained activity (15%).

What is deliberately not generated

Contacts and Tech Stack panels are empty by choice. Filling them would mean inventing contacts and technology for real companies from news headlines, and you'd repeat that on a call. Each panel names the real data source (e.g. ZoomInfo, Lusha, BuiltWith) that would populate it instead of guessing.

Hosting

Runs on Vercel by default. server/index.mjs (zero-dependency), the Dockerfile, and npm run ingest are everything needed to run it anywhere else — see HOSTING.md for self-hosting and TLS. There's no login screen: anyone with the URL can open the dashboard, by design — this is a single-user tool with no sensitive data behind it.

Data pipeline

Signal data lives in Postgres (Neon, connected via Vercel's native integration), not in the JS bundle.

• GET /api/signals — the frontend fetches this at runtime.
• POST /api/ingest — pulls Google News RSS for every company on your tracked list (api/_lib/watchlist.js), dedupes against existing rows, normalizes new items via a Claude API call (api/_lib/normalize.js), matches them against your tracked companies (api/_lib/matchCompanies.js), and inserts them. Triggered on a daily cron (vercel.json), protected by CRON_SECRET.
• db/migrations/ — schema, applied in order.
• db/backfill-relevance.mjs — one-time/re-run relevance scoring pass, useful after editing the rubric.
• POST /api/reviews — durable "Mark Reviewed" state.
• GET /api/ingest-status — the most recent ingest run's outcome, so the frontend can show "last synced" honestly.
• GET/POST/DELETE /api/companies — the tracked-companies list: add, update, or stop tracking a company. This is the entire configuration surface of the app.

Row shape (camelCase over the wire, snake_case in Postgres):

{
  "id": "string",
  "headline": "string",
  "summary": "string, 2-3 sentences",
  "sourceUrl": "https://...",
  "date": "YYYY-MM-DD",
  "scope": "macro | micro",
  "entity": "string",
  "signalType": "see shared/signalTypes.js for the canonical list, or 'community insight' for last30days-derived rows",
  "origin": "seed | news | community",
  "outreachRelevance": "integer 1-5, or null if not yet scored",
  "reviewed": "boolean",
  "matchedCompanies": ["companies on your tracked list this signal names"]
}

Tracked companies (the thing that drives everything):

{
  "companyKey": "string, normalized",
  "companyName": "string",
  "status": "customer | prospect | null",
  "isCompetitor": "boolean",
  "note": "string or null"
}

One-time setup (required before this works)

1. Fork or clone this repo, and deploy it to Vercel (connect your GitHub account, import the repo, deploy — Vercel's free tier covers this comfortably).
2. Connect Postgres: Vercel dashboard → your project → Storage → Connect Database → choose Neon (or Vercel Marketplace → Neon) → Connect to Project. This injects DATABASE_URL automatically.
3. Set environment variables (Vercel dashboard → Project → Settings → Environment Variables):
   • ANTHROPIC_API_KEY — get one at console.anthropic.com. Used by the ingest pipeline's normalization step; capped at $10/month (see below), so this can't run away on you.
   • CRON_SECRET — any random string you generate; Vercel automatically sends it as Authorization: Bearer $CRON_SECRET when invoking the cron job.
4. Run the migrations: open the Neon/Postgres query editor in the Vercel dashboard and run each file in db/migrations/ in order.
5. Add your first company: open the deployed app, go to My Accounts, click "Track a Company." The next scheduled ingest run will pick it up automatically.
6. Verify ingestion manually before trusting the cron: curl -H "Authorization: Bearer $CRON_SECRET" https://<your-deploy>/api/ingest and check the response summary (queried, rawItems, afterDedupe, normalized, inserted, companyMatched, errors, fetchErrors). A non-empty fetchErrors means one of your search terms couldn't be searched at all (network issue) — worth a look if rawItems stays at 0 for a company you know is in the news.

On watchlist size: each run queries the news once per company you're tracking, plus anything you've added to STANDING_WATCHLIST in api/_lib/watchlist.js (empty by default — add your own industry/region searches there if you want them). api/_lib/fetchNews.js's fetch concurrency is 10. A very large tracked list (100+ companies) is capped per run — see the comment in watchlist.js.

Monthly Claude API budget guardrail

api/_lib/budget.js enforces a hard $10/calendar-month ceiling on Claude API spend from the ingest normalizer, independent of how often /api/ingest runs. Every normalization call's real token usage is logged to the llm_spend table immediately after the call; before each call, normalizeItem sums this month's spend and stops calling the API once at or above the cap, reporting summary.budgetExceeded: true instead of silently truncating the feed. To change the cap, edit MONTHLY_BUDGET_USD in api/_lib/budget.js.

Community signals (optional, advanced)

A third data source beyond the news pipeline: the last30days Claude Code skill run locally against tracked entities, transformed into the same schema, and pushed in as origin: 'community' rows. This cannot run on Vercel — it needs a real Claude Code session with genuine internet access, which a serverless function doesn't have. See scripts/local/README.md for setup. Entirely optional; skip it if you just want the news pipeline.

Development

npm install
npm run dev

npm run dev only runs the Vite frontend — /api/* routes need vercel dev (or a deployed environment) to actually execute. Without that, the UI falls back to a small fixture of clearly-fake example data (src/data/data.json) so you can see the layout before connecting a real database.

Build

npm run build

Outputs a static site to dist/, ready to deploy on Vercel.

License

MIT — see LICENSE. Free to use, modify, and redistribute.