News Intelligence Agent

Ask any news question in plain English. Get back a cited, multi-source, multi-language report.

This Actor reads the news for you and tells you what matters. You type a question, it does the research, you read the answer.

What it does

You give it:

• A question like "What is happening with oil prices this week?"
• Your Anthropic Claude API key

It gives you:

• A short summary of the top findings
• The 5 most important headlines
• 3 to 7 themes grouping related articles together
• Any surprises or anomalies flagged
• A list of the people, companies, and countries most mentioned
• The overall mood (positive, neutral, negative, or mixed)
• 3 follow-up questions you might want to ask next
• A full markdown report ready to read or share
• The raw articles with summaries, keywords, and impact scores

All in the language you pick (English, Arabic, Chinese, Spanish, and 10 more).

Who it is for

• Government and policy teams tracking global events
• Analysts who need a digest before a meeting
• PR and brand teams monitoring coverage
• Investor research on signals across markets
• Journalists tracking patterns across many sources
• Corporate strategy teams watching competitors and regulation

How it works in 30 seconds

1. You type your question in plain English.
2. Claude reads the question, picks a balanced source set, and writes 3 to 8 search queries that cover the angle.
3. Apify discovers articles via Google News (with rag-web-browser as a fallback) and routes each URL to the right extractor: dedicated specialists for Bloomberg, Reuters, NYT, FT, Guardian, Al Jazeera; Smart Extractor for everything else.
4. Claude reads each article, decides if it is relevant, scores its impact, and pulls keywords, entities, and quotes.
5. Claude looks across all the articles, groups them into themes, flags anything surprising, and writes you a final brief.
6. You get the output as a structured dataset and a markdown report.

The whole thing takes 2 to 5 minutes. You do not have to pick sources, write search queries, or read 50 articles yourself. Claude does that.

Quick start

1. Get an Anthropic API key at console.anthropic.com (starts with sk-ant-).
2. Open this Actor in the Apify Console.
3. Type your question in the Your question field.
4. Paste your Anthropic key in the Anthropic Claude API key field.
5. Paste your personal Apify API token in the "Apify API token" input field. Find it at Apify Console -> Settings -> Integrations -> Personal API tokens.
6. Click Start.

That is the entire setup. Every other field has a sensible default.

The Apify token is required because the runtime token issued to the Actor is scoped down and cannot call the child scrapers. Your personal token is used only during the run and is never reused elsewhere.

Example questions

• What global events this week could materially affect Dubai's economy? Focus on oil, trade, tourism.
• How is the new EU AI regulation being covered across European and American press?
• What are competitors saying about supply chain disruption in semiconductors?
• Track misinformation patterns around the latest election in Country X.
• What is happening with central bank rate decisions in the last 7 days?

Input fields

You only need to fill in three fields. Everything else is optional.

Field	What it does	Default
Your question	Plain English. Be specific.	required
Anthropic API key	Starts with sk-ant-. Apify never stores it.	required
Apify API token	Used to call the child scrapers. Find at Apify Console -> Settings -> Integrations.	required
Use direct RSS feeds	Pulls articles directly from major publishers' RSS feeds in real time, removing the Google News indexing lag.	on
Include Reddit	Adds Reddit posts as a community sentiment signal.	off
Include Twitter / X	Adds tweets from verified accounts as a breaking-news signal.	off
Wikipedia enrichment	Fetches Wikipedia summaries for top entities so Claude can ground its analysis in factual context.	off
Claude model	Sonnet 4.6 is the default. Haiku is cheap and fast. Opus is thorough and slow.	Sonnet 4.6
Output language	What language you want the final report in. 14 options.	English
Source selection mode	Auto lets Claude pick. Manual uses only your sources. Hybrid combines both.	Auto
Manual source list	Domains like bbc.com, reuters.com. Only used in manual or hybrid.	empty
Excluded sources	Domains to skip.	empty
Semantic tags	Extra keywords Claude should weigh higher. Example: OPEC, NATO.	empty
Country focus	ISO codes like US, CN, RU, SA, AE. Leave empty for global.	empty
Translate non-English articles	If on, Claude reads sources in any language. If off, only output-language sources.	on
Time window	Last 24 hours, 7 days, or 30 days.	7 days
Research depth	Shallow (20 articles), medium (50), or deep (100).	medium
Max articles cap	Hard ceiling on articles analyzed.	50
Bypass paywalls	Uses Googlebot user-agent to access Bloomberg, FT, WSJ.	on
Custom instructions	Power users only. Extra guidance for Claude.	empty
Max Claude spend	Hard ceiling on token cost. The run stops cleanly if exceeded.	$5
Analysis features	Toggles for keyword extraction, sentiment, entities, quotes, etc.	all on
Webhook URL	POSTs the markdown report to Slack, email, or any URL.	empty

What you get back

One record in your default Apify Dataset:

{
  "userQuery": "your question",
  "outputLanguage": "English",
  "generatedAt": "2026-05-12T...",
  "executiveSummary": "3 to 5 bullet points",
  "topHeadlines": [{ "title": "...", "url": "...", "impactScore": 9, "oneLineWhy": "..." }],
  "themes": [{ "name": "...", "summary": "...", "articleUrls": [], "averageImpactScore": 7 }],
  "anomalies": [{ "headline": "...", "url": "...", "impactScore": 9, "why": "..." }],
  "keyActors": [{ "name": "OPEC", "type": "organization", "mentionCount": 12, "role": "..." }],
  "overallSentiment": { "summary": "negative", "breakdown": { ... }, "explanation": "..." },
  "relatedQuestions": ["follow-up 1", "follow-up 2", "follow-up 3"],
  "confidenceScore": 0.82,
  "confidenceReason": "Good coverage, recent sources, consistent themes",
  "articles": [
    {
      "title": "...",
      "url": "...",
      "source": "reuters.com",
      "publishedAt": "...",
      "language": "en",
      "summary": "...",
      "keywords": [],
      "entities": { "organizations": [], "countries": [], "people": [] },
      "sentiment": "negative",
      "impactScore": 8,
      "anomalyFlag": false,
      "quotes": []
    }
  ],
  "sourceCoverage": { "totalArticlesIncluded": 38, "uniqueSources": 14, "sources": [], "languageBreakdown": { "en": 28, "ar": 6, "zh": 4 } },
  "markdownReport": "the full readable brief in markdown",
  "costEstimateUsd": { "claudeTokens": 0.87, "totalSpentSoFar": 0.87 },
  "tokensUsed": { "input": 95000, "output": 12000 }
}

The markdown report has 8 sections in order: Executive Summary, Top Headlines, Themes, Anomalies, Key Actors, Overall Sentiment, Related Questions, Article Index.

Cost

Pricing for this Actor is not finalized yet. It is in early access. The cost model below reflects only your underlying compute and Claude API spend during the beta period.

Two underlying bills today:

1. Apify platform compute, billed on your Apify account. Roughly $0.30 to $0.50 per typical run.
2. Anthropic Claude tokens, billed directly by Anthropic. Set your spending limit per run in the input. Defaults to $5 ceiling.

Typical run with Claude Sonnet 4.6:

• Shallow (20 articles): about $0.30 in Claude tokens
• Medium (50 articles): about $1.00 in Claude tokens
• Deep (100 articles): about $2.50 in Claude tokens

Claude Haiku 4.5 is 8 to 10 times cheaper. Claude Opus 4.7 is 5 times more expensive.

You pay Anthropic directly. The Actor does not mark up your Claude usage during the beta.

A pay-per-event price model will be added once the Actor leaves beta. Expected components:

• A small per-run start fee
• A per-article fee for each article that passes the relevance filter
• A per-digest fee for the final synthesized report

Final numbers and per-customer-tier discounts will be published before launch. If you want to lock in early-access pricing as a design partner, contact the maintainer via the Issues tab.

Limits

• Some sites block automated access even with paywall bypass on. Document any consistent failures and we will look at it.
• Non-Latin language quality depends on Claude's training. Tested broadly on Arabic, Chinese, Russian, Japanese, Spanish, German, French.
• Runs are not real-time. Each run pulls fresh articles within the chosen time window.
• If your Anthropic balance is exhausted mid-run, the run fails cleanly with a clear message. Set spending limits in your Anthropic console.

Tech stack under the hood

Discovery (Stage 2). Multi-channel, toggle what you need:

• data_xplorer/google-news-scraper-fast as the primary discovery channel. Native news date filtering, returns decoded original article URLs. Always on.
• apify/rag-web-browser as a fallback when Google News returns thin results for a query.
• Direct RSS feeds (automation-lab/rss-feed-reader) from Reuters, BBC, AP, Bloomberg, FT, Guardian, NYT, WaPo, AlJazeera, CNN, CNBC, Politico. Real-time, removes Google News indexing lag. On by default.
• Reddit (trudax/reddit-scraper-lite). Opt-in via enableReddit. Community signal, sentiment.
• Twitter / X (apidojo/twitter-scraper-lite). Opt-in via enableTwitter. Breaking news 2-6 hours ahead of mainstream press.
• Wikipedia entity enrichment (automation-lab/wikipedia-scraper). Opt-in via enableWikipedia. Fetches 1-paragraph summaries for the most-mentioned organizations, people, and countries so Claude can ground its analysis in factual entity context.

Source-tier weighting in synthesis:

• Each article is automatically tagged Tier 1 (wire services and major publishers), Tier 2 (established regional and topic specialists), or Tier 3 (niche or blog).
• Stage 5 weights Tier 1 claims higher in corroboration. Social signal (Reddit, Twitter) is labelled "[social signal]" rather than "[single-source]" so the reader knows it's a leading indicator, not a confirmation.

Extraction (Stage 3): domain-aware routing.

Tier-1 publishers go to dedicated Store Actors that handle paywalls and bot defences. Cheerio-based generic extractors get blocked on these sites, so each gets its own specialist:

• nytimes.com -> xtracto/nytimes-scraper
• bloomberg.com -> xtracto/bloomberg-news-article-scraper
• reuters.com -> xtracto/reuters-scraper
• ft.com -> xtracto/ft-scraper
• theguardian.com -> xtracto/guardian-scraper
• aljazeera.com -> piotrv1001/aljazeera-scraper
• bbc.com / bbc.co.uk -> nexgendata/bbc-news-scraper

Everything else falls through to lukaskrivka/article-extractor-smart with apify/website-content-crawler as a last-resort fallback. Any URL that returns empty from a dedicated extractor is automatically retried through the Smart Extractor, so coverage never drops silently.

Brain (Stages 1, 4, 5):

• Anthropic Claude API for planning, per-article analysis, and final synthesis.

Three Claude pass groups per run: one to plan, one per article batch to analyze, one to synthesize. That is what makes this an agent, not just a wrapper.

Troubleshooting

If a run finishes with fewer articles than expected, check the run log for these patterns:

• "exceed the memory limit" on a dedicated extractor. You are hitting the account-wide memory cap (8GB on Starter, 32GB on Scale). The Actor lowers parallelism and memory automatically in v0.2.4, but a heavy parallel load on your account from other Actors can still push you over. Wait for other runs to finish, or upgrade your Apify plan.
• "Insufficient permissions" or "LIMITED_PERMISSIONS" on a child Actor. Your runtime token is scoped down. Two fixes: pass your personal Apify API token in the apifyToken input field, or restart the run from the Console UI and grant "Call other Actors" in the Run options dialog.
• "Request blocked - received 401 / 403" inside the rag-web-browser logs. This is expected for paywalled / cookie-walled sites. rag-web-browser still returns the URL even when it cannot fetch the content, so the URL is handed off to the right dedicated extractor in Stage 3.

Support

Use the Issues tab on this Actor page to reach the maintainer.