News API for AI agents
If your agent needs to know what is happening in the world, a news API that returns articles hands it the wrong unit. CLSTR returns events and situations: one record per happening, with the history attached, over REST and MCP.
The problem: twenty rewrites, no memory
When something happens, dozens of outlets write it up. A keyword search over articles gives your agent twenty rewrites of one event, each under a slightly different headline. The agent spends its context window merging duplicates, and still cannot tell a new development from a recycled one.
Ask again tomorrow and it starts from zero. An article feed has no memory of how a situation got to where it is.
What an agent actually needs
- One event per happening. Twenty outlets covering the same thing should arrive as one item, with a count of how many are covering it.
- The situation over time. Related events linked into one thread with a maintained summary and a timeline, so the agent can answer "what has happened here over the last six weeks", not only "what is the latest".
- Structured output. Fields it can read (title, summary, source count, dates, countries, significance) instead of prose it has to parse.
- Canonical links. A stable URL for every event and every situation, so an answer can link to where it came from and a reader can open the sources.
How CLSTR answers it
CLSTR reads about 107,000 articles a day, from more than 75,000 sources over the 30 days to October 21, 2026, and groups coverage of the same event into a cluster: deduplicated, with linked sources and the outlet count, and a significance score from 1 to 10. Related clusters link into a situation, which carries a maintained summary and a full timeline.
- REST and MCP, same data. A versioned JSON API at
https://api.clstr.news/v1and a remote MCP server athttps://mcp.clstr.news/mcp, on the same key and the same caps. - Four tools.
get_top_situations,search_situations,get_situation_timelineandget_cluster. When to call each one is on the MCP server for news page. - OAuth for assistants. Claude and ChatGPT connect with a sign-in, not a pasted key. Frameworks and CLIs use a Bearer key.
- Keyless taster. get_top_situations and get_situation_timeline answer with no key at all, 3 calls per day per network, so you can point a client at it before you sign up.
- Free key, no card. 100 requests a day plus search. Caps are hard limits: past the cap you get a 429, never a bill. Paid tiers are on the pricing page.
How events are formed
- Multi-outlet by construction. An event surfaces once more than one outlet is covering it, so your agent gets stories the press is reporting, not single-source noise, and the outlet count shows how broadly.
- It shows what the press reports. A summary says what the coverage says, with the sources linked.
- English output, many languages in. Titles and summaries are written in English. The sources span 58 languages over the 30 days to October 21, 2026, and English is at most 29% of what CLSTR reads (as of September 2026, measured on the observatory).
- History windows. Listing and search reach back 7 days on Free and 30 days on paid tiers. A situation's timeline goes back to where it started.
How it compares
Against the generic alternatives, on the dimensions an agent cares about. These are categories, not named products, and individual products vary.
| CLSTR | Raw news APIs | Web search tools | RSS feeds | |
|---|---|---|---|---|
| Deduplication | One cluster per event, with a source count | Usually one row per article; merging is your job | Ranked pages; one event can fill the results | One item per article, per feed |
| Memory over time | Situations with a timeline across weeks | You build it | None; each query starts fresh | Only the recent items the feed holds |
| Structured output | JSON over REST; MCP tools declare an outputSchema | JSON articles | Varies; often text snippets | XML items, no event structure |
| Auth for assistants | OAuth sign-in for Claude and ChatGPT; Bearer key for frameworks | Usually an API key you manage | Built into some assistants | None needed; no connector |
| How widely reported | Outlet count on every event | Count it yourself | Varies | One feed at a time |
Start
- Connect Claude, ChatGPT or a framework: MCP server for news.
- Read the endpoints, fields and caps: API and MCP reference.
- Compare the tiers: pricing.
- See what the pipeline reads, measured: the observatory.