The shape of the problem

Ask a model about yesterday's election result or this week's product launch and it will either say it doesn't know or, worse, guess. The fix is retrieval: fetch recent articles first, hand them to the model, and ask it to answer from those articles only.

A news API is the fetching part. You send a query and filters, and get back a list of articles as structured data.

What a news API returns

  • A title, a link, the publisher, a publication time and usually a short description.
  • Sometimes the full text, sometimes only a snippet. This is the biggest difference between services.
  • Metadata such as language, country, category or detected topics.
  • Paging information, so you can ask for the next batch.

The parameters that matter

Most services share the same core: a search query, a date range, language and country filters, a list of sources to include or exclude, and a sort order. Learn the query syntax early. Exact phrases, AND, OR and NOT make the difference between fifty relevant articles and five thousand loosely related ones.

How to choose one

  • Coverage: how many publishers, which countries and which languages. A big number matters less than whether your market is in it.
  • Freshness: how many minutes pass between publication and availability.
  • History: whether you can search last year, or only the last month.
  • Full text or snippets, and what you are licensed to do with either.
  • Deduplication: the same wire story appears on hundreds of sites.
  • Limits and price: requests per day, results per request, and what happens when you exceed them.
  • Reliability: an uptime record and a status page.

Free and open options

You don't have to start with a paid plan. RSS and Atom feeds are still published by most news sites and need no key at all. The GDELT Project offers an open, global index of news coverage that researchers have used for years. Several newspapers run their own developer APIs with free tiers, and general news aggregator APIs usually have a free plan for development.

Free tiers typically limit history, delay results or forbid commercial use, so read the terms before you build a product on one.

Giving a model the news through tools

Instead of fetching articles yourself and pasting them into a prompt, you can describe the API to the model as a tool and let it decide when to call it. The Model Context Protocol (MCP) has become a common way to package such tools, and many data providers now publish an MCP server alongside their REST API.

Keep tool results small. Return titles, dates, sources and short summaries first, and fetch full text only for the two or three articles the model selects.

Using the results well

  • Tell the model to answer only from the supplied articles and to say so when they don't cover the question.
  • Pass publication dates and ask the model to state them. "Recent" means nothing without a date.
  • Ask for citations by link, and show them to the user.
  • Prefer two independent sources for any important claim.
  • Watch your token count: ten full articles can cost more than the answer is worth.

What these datasets get wrong

Coverage is uneven: English and large countries dominate. Paywalled sites are often missing or truncated. Syndicated copies inflate counts, so "a thousand articles" may be thirty stories. And volume of coverage measures attention, not truth. A rumour repeated five hundred times is still a rumour.

Copyright and terms

This is general information, not legal advice. Headlines and links are generally safe to show. Republishing full articles is not, unless your licence says so. Summarising for a user and linking to the source is the common pattern. Read the API's terms for what you may store and for how long.

If you don't want to build anything

Most people don't need an API. They need an assistant that can look things up. In UncutGPT, supported models can search the web when you turn web search on, and they answer with what they found. For frank questions about the news, including ones mainstream assistants sidestep, pair web search with an uncensored model and ask for the strongest version of each side.