Limits & pricing
Caps are hard limits: requests past the cap get a 429, and there are no overage charges, ever. Search is capped separately because each search runs an embedding.
| Tier | Price | Requests/day | Requests/min | Searches/day | History window | Capabilities |
| Anonymous (MCP taster) | No key | 3 | 3 | Keyed only | 7 days | Preview summaries in lists |
| Free | Free | 100 | 20 | 10 | 7 days | Preview summaries in lists |
| Builder | $29/mo | 5,000 | 60 | 250 | 30 days | Full summaries in lists |
| Scale | $299/mo | 100,000 | 1,000 | 10,000 | 30 days | Full summaries in lists |
The anonymous row is the MCP taster: 2 tools (get_top_situations and get_situation_timeline), no signup, counted per network rather than per person and capped platform-wide. Past 3 calls a day, an anonymous call returns a link to a free key instead of data. Everything else, including search, needs a key.
Caps apply per account, not per key: an account may hold up to 5 active keys and they all draw on the same limits. The history window is the widest days value a tier can ask for; larger values are clamped, not rejected. Search reaches back at most 30 days, and never beyond the tier's history window.
Builder and Scale are self-serve: upgrade on the pricing page. If the caps still do not fit, ask for a custom quote on the contact page.
You may display the output to your users, provided you keep the links to the original sources wherever that output is shown (see the Terms).
Every authenticated response carries X-RateLimit-Limit-Day, X-RateLimit-Remaining-Day and X-RateLimit-Limit-Minute (all exposed to browsers via CORS). There is no remaining-count for the per-minute window: it is enforced per Cloudflare location and reports no count, so any number here would be a guess. Pace against the daily remaining, and back off on a 429 using Retry-After. /search additionally carries X-RateLimit-Limit-Search and X-RateLimit-Remaining-Search for the daily search cap, which is the limit a free key reaches first. Responses rejected before authentication (a missing or invalid key) carry none of them, because there is no key to report usage for. There is also a generous per-address ceiling in front of authentication; a normal client reaches its own tier cap long before it.
A 429 carries Retry-After in seconds: the exact time left in the window that fired, not a fixed guess.
MCP server
Your agent can search the web. It can't tell you what's happening: which developing situations matter, what 40 sources collectively report, and how a story evolved over six weeks. CLSTR can, over MCP.
One event becomes a single cluster, grouped from the outlets covering it and deduplicated, with a source count and a significance score. Related clusters link into a situation that carries a maintained summary and a full timeline, so an agent gets memory, not just the latest headline.
Remote server, no install. Same API key, same caps (one tool call counts as one request).
Endpoint: https://mcp.clstr.news/mcp
Auth: OAuth sign-in (assistant connectors) · Bearer API key (configs, CLIs, frameworks) · keyless taster (try it)
Tools: get_top_situations · search_situations · get_situation_timeline · get_cluster
Assistants (Claude, ChatGPT): sign in, do not paste a key. Add CLSTR as a connector with the URL https://mcp.clstr.news/mcp and sign in when prompted. A free CLSTR account is enough, no card. The connector runs an OAuth sign-in against your account, so it gets all four tools and your plan's caps with no header to manage. Disconnect it any time under Connected apps in Settings.
Choose this lane before you add the connector: a connector added without a sign-in cannot be switched to OAuth afterwards, only removed and added again. Use the full URL including /mcp.
Frameworks, CLIs and clients with an mcpServers config: API key. Anything that keeps its servers in a config file, and any client with no OAuth flow of its own, authenticates with a Bearer header.
"clstr": {
"url": "https://mcp.clstr.news/mcp",
"headers": { "Authorization": "Bearer clstr_YOUR_KEY" }
}
stdio-only client? Bridge with npx mcp-remote https://mcp.clstr.news/mcp --header "Authorization: Bearer clstr_YOUR_KEY", which is the same key, not a sign-in.
Keyless taster, to try it before you sign up: for curl, demos, and frameworks that cannot hold a credential. get_top_situations and get_situation_timeline answer anonymously, 3 calls per day per network (burst 3 per minute), and anonymous timelines come back at most 50 entries deep. After that, a call returns a link to a free key instead of data. A free key lifts that to 100 requests per day, full timeline depth, and search.
If you are setting up an assistant connector, sign in instead: a keyless connector cannot be upgraded later without removing it and adding it again.
Check it works, no key:
curl -s https://mcp.clstr.news/mcp -H "Content-Type: application/json" \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
"params":{"name":"get_top_situations","arguments":{"window":"24h","limit":3}}}'
The tools obey the same windows as the REST endpoints: search_situations reaches back at most 30 days, and get_top_situations takes 24h, 7d or 30d. Refusals a caller can act on (a cap, an unknown id) come back as readable tool results; transport problems come back as JSON-RPC errors with the HTTP status to match.
get_top_situations takes the same filters as GET /situations: category, country (comma separated ISO codes, up to 10) and sort (relevance or recent). One difference from REST: the tool rejects a country code or a sort value it cannot use instead of dropping it, so an agent is never told it filtered by something that was ignored.
Every tool declares an outputSchema and returns structuredContent next to its markdown, on protocol revision 2025-06-18. That payload is the same object the REST equivalent in the table above publishes, with two differences. First, clstr.news links in it carry ?ref=mcp, exactly as the markdown links do, so the channel is measurable. Publisher links on articles are never touched. Second, each search_situations row also carries a situation block (id, title, slug, cluster_count; null when the event belongs to no situation) that GET /search rows do not, because the tool groups its markdown by situation. A client can read fields instead of parsing text; the markdown is unchanged and still comes back for clients that do not read structured results.
Both list tools report summary_mode in structuredContent, beside next_cursor: the mode actually served, which on get_top_situations is preview unless the account is on Builder or Scale. The downgrade is never an error; the tool returns the preview list and adds one line saying which call gives you one situation's summary in full. get_top_situations also reports window_days: the window actually served, in days, after your tier's history window.
get_top_situations and search_situations take summary: "full", and get_cluster takes articles (1 to 100), mirroring the query parameters on their REST twins. Another difference from REST is scoped to those two parameters: summary and articles reject a value outside the declared range with a JSON-RPC -32602 rather than clamping it, so an agent is never told it asked for something it did not get. limit, days and timeline_limit still clamp, as they always have.
get_top_situations and search_situations paginate: when more results exist, the tool ends with a cursor to pass back. For get_top_situations, send the same filters with it: a cursor resumes only under the window, category, country and sort that produced it. For search_situations, each page re-runs the search, so one page counts as one search against the cap.