Deep-research agents get their answers by searching, reading, and searching again. The router asks eight indexes at once, so your agent gets the whole picture in one call and spends its tokens thinking instead of retrying.
For Claude, Codex, Pi, Openclaw, Cursor, and other coding agents.
The router asks eight indexes at once, so your agent gets the whole picture in one call and spends its tokens thinking instead of retrying.
01 / The problem
Every search API has blind spots, and they don't overlap. Exa's neural index, Brave's crawl, Parallel's, You.com's: each one is built for something different and sees a different web. Wire your agent to one of them and it inherits that index's gaps on every question it ever asks.
Rephrasing doesn't fix that. A second opinion does. So does an eighth.
02 / What changes
One call fans out to every index in your tier, in parallel. Results come back merged, de-duplicated and labelled by source. No index gets to be the only one that mattered.
When the first result set is thin, a single-index agent searches again, and every extra turn costs model tokens. Fan-out front-loads the coverage so your agent has less reason to go back.
No more juggling accounts, rate limits and invoices. The router holds the provider keys. You hold one.
Every response records which index returned what, how fast, and what it cost. Your agent's research trail becomes data: inspectable, replayable, yours.
03 / How it works
POST /v1/search with the question. Pin a tier, a provider list or a domain filter if you want.
Every index in your tier runs at the same time, under a hard per-query cost ceiling. If one fails, the rest still land.
Pooled, de-duplicated by URL, ordered. Titles, links and excerpts, ready for a model to read.
Per-index status, latency and cost, kept against the request id.
04 / Built for agents
The drop-in miner installs as a skill with permissions pre-set. No daemon, no MCP server, nothing running in the background.
/dropin-miner off — when you don't want it
/dropin-miner on — when you do
The router gives your model the sources, not a paragraph written by someone else's model. Your agent reads the web, not a summary of it.
Router traffic takes part in Twilight's proof-of-work mining. Agents that search, earn.
How mining works →05 / Benchmark
We benchmark on BrowseComp with OpenRouter's own prompts and graders, unmodified, and we publish every run, including the ones we lost. The harness, the methodology and the transcripts are in the open.
On BrowseComp, one fan-out search reached what single-index agents needed N searches to reach.See the benchmark →
curl https://router-api.nyks.dev/v1/search \
-H "Authorization: Bearer $KEY" \
-d '{"query": "your question here"}'