Web search (Keenable)

Search the web and read pages straight from your code — with the same key you use for models. Two ways to call it: as a regular model through /v1/chat/completions, or as dedicated endpoints returning plain JSON.

Pricing

$0.005 per request — that is $5 per 1,000 requests.

The price is per call: one call is one request, no matter how many results come back. Finding 50 sources costs the same as finding one.

Failed requests are not billed — neither a search failure nor bad parameters.

what	counts as
a search returning 1 result	1 request
a search returning 50 results	1 request
reading a page	1 request
any error	0

Way 1 — as a model

Works in Cursor, Claude Code, OpenCode and any OpenAI client with no changes at all: just pick the model.

Models:

model	what it does
keenable/search	searches the web, returns a list of sources with snippets
keenable/read	reads a page and returns its text as markdown

Search

curl https://nordrouter.com/v1/chat/completions \
  -H "Authorization: Bearer $NORDROUTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "keenable/search",
    "messages": [{"role": "user", "content": "bitcoin price today"}]
  }'

You get a normal assistant message listing the sources: title, publication date, full URL and a snippet for each.

Reading a page

Put the URL first in your message. After it you may add what exactly you are looking for on that page.

curl https://nordrouter.com/v1/chat/completions \
  -H "Authorization: Bearer $NORDROUTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "keenable/read",
    "messages": [{"role": "user", "content": "https://en.wikipedia.org/wiki/Tokyo population figures"}]
  }'

Without the extra instruction you get the whole page as markdown. With it you get only what you asked for — which is both faster and noticeably cheaper in tokens on your side when the result feeds into a model.

Why bother with the instruction

A Wikipedia article is tens of thousands of characters. If you need one number, the instruction saves both time and the tokens of your next step.

Python

from openai import OpenAI

client = OpenAI(
    api_key="your-key",
    base_url="https://nordrouter.com/v1",
)

found = client.chat.completions.create(
    model="keenable/search",
    messages=[{"role": "user", "content": "AI news this week"}],
)
print(found.choices[0].message.content)

Way 2 — dedicated endpoints

When you need results with fields, dates and filters rather than prose, take the JSON.

POST /v1/search

curl https://nordrouter.com/v1/search \
  -H "Authorization: Bearer $NORDROUTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "neural network news",
    "max_results": 5,
    "published_after": "7d"
  }'

Parameters:

parameter	type	default	what it does
query	string	required	what to search for
site	string	—	search only this domain, e.g. github.com
max_results	number	10	how many results, 1 to 50
snippet_max_chars	number	—	snippet length, 180 to 10,000 characters
published_after	string	—	material published after this moment
published_before	string	—	published before it
acquired_after	string	—	added to the index after
acquired_before	string	—	added to the index before
query_time	string	—	search the index as it stood at this moment

Dates accept three forms: 2026-08-01, a full ISO 8601 timestamp, or a relative delta — 30min, 12h, 7d, 3mo, 1y.

Published and indexed are not the same

published_* filters by the date of the material itself, acquired_* by the date the page entered the index. The difference matters: a page can be rewritten without touching its publication date. If you want "what appeared online in the last day", use acquired_after.

Response:

{
  "query": "neural network news",
  "count": 5,
  "results": [
    {
      "title": "Article headline",
      "url": "https://example.com/article",
      "description": "Short description",
      "snippet": "An excerpt from the page text…",
      "published_at": "2026-08-25T10:30:00Z",
      "acquired_at": "2026-08-25T11:02:00Z"
    }
  ]
}

GET /v1/fetch

curl -G https://nordrouter.com/v1/fetch \
  -H "Authorization: Bearer $NORDROUTER_KEY" \
  --data-urlencode "url=https://example.com/article" \
  --data-urlencode "prompt=extract the pricing table"

Parameters:

parameter	type	default	what it does
url	string	required	the page address
max_chars	number	50,000	how many characters to return
prompt	string	—	what exactly to extract, up to 2,000 characters
live	true / false	false	fetch from the source instead of the index

Response:

{
  "url": "https://example.com/article",
  "title": "Headline",
  "description": "Description",
  "author": "Author",
  "content": "# Page text as markdown…",
  "published_at": 1787777084
}

When you need live

The index answers faster, but it does not hold everything. If the page is fresh or closed to crawlers, set live=true and it will be read straight from the site. Same price.

Limits

rate	2 requests per second per key
snippet length	180 to 10,000 characters
results per call	up to 50
extraction instruction	up to 2,000 characters

Going over the rate returns 429 with a Retry-After header — retry in a second.

Errors

code	what it means	what to do
400	query or url missing, or the address has no http://	fix the request
401	key not accepted	check Authorization
402	balance exhausted	top up in the dashboard
429	too fast	retry in a second
502	search did not answer	retry; nothing was billed
503	search temporarily unavailable	we are on it; nothing was billed

FAQ

Am I billed for an empty result set? Yes, if the search ran and honestly answered "nothing found" — the work was done. Only errors are free.

Can I stream it? No, and there is no need: results arrive in one piece within a fraction of a second.

Are tokens counted? The price is per call, tokens do not affect it. They are still reported in the model response so client libraries do not treat it as empty.

How is this different from the search/… models? Those search first and then answer with a language model — you pay for two jobs. Here you get the raw results and decide what to do with them yourself.