Research Desk
No subscription — pay in Ounie credits

Every claim
names its source.

Ask a question. Research Desk searches the live web, reads the pages, and writes you a brief in which every sentence carries the number of a source it actually fetched. Anything it could not attribute is dropped before you see it — and the count of what was dropped ships with the brief.

Run your first briefWire it into an agent
From
15 cr
one brief
Sources
up to 24
read in full
Invented cites
0
by construction
KEY FINDINGSnature.com1reuters.com2arxiv.org3sec.gov41234
The depth ladder

Depth buys work, not adjectives.

Three rungs, and each one changes what actually happens upstream: how many angles the question is searched from, how many pages come back, and how much of each page is read. The deep rung adds a second pass that goes looking for whatever the first pass admitted it could not answer.

Quick

15 cr

One search, headlines read. For a fast orientation on a question you already half-know.

READ PER SOURCE1,200 characters
Search angles
1
Sources read
up to 6
Gap-closing pass
no
Typically takes
~25s
Default

Standard

35 cr

The question is broken into three angles and each source is read properly. The default.

READ PER SOURCE3,000 characters
Search angles
3
Sources read
up to 14
Gap-closing pass
no
Typically takes
~60s

Deep

75 cr

Six angles, long reads, then a second pass that searches the gaps the first pass left open.

READ PER SOURCE6,000 charactersthen searches its own gaps
Search angles
6
Sources read
up to 24
Gap-closing pass
yes
Typically takes
~150s
The citation contract

A claim that cites nothing real does not ship.

The model is asked for claims with source numbers on them. Then those numbers are checked, one at a time, against the pages we really fetched. A number that points at nothing is thrown away; a claim left holding nothing but thrown-away numbers is thrown away with it. Nothing is invented to fill the gap.

WHAT THE MODEL WROTEWHAT SHIPSCHECK1414292source 9 was never fetched7cited nothing real — dropped2 citations rejected · 1 claim dropped — both counted in the brief you receive

Surviving sources are renumbered so that 1 is the first thing you are asked to check, and a page nobody cited is reported in the stats rather than padded into the list. Every source arrives with a verbatim quote, so checking a claim is a glance, not a click-through.

For developers & AI agents

“Go research X” is one call away.

Research is the subtask an agent delegates most and verifies least. Every brief here comes back with its sources attached, so the calling agent can check a claim before it acts on one — and so can the person reading its output afterwards.

REST

POST a question, get a brief id back, poll it. Bearer key, JSON in, cited JSON out.

curl -X POST https://research.ounie.com/api/briefs \
  -H "Authorization: Bearer rsd_live_…" \
  -d '{"question":"…","depth":"standard"}'
MCP

Native tools for Claude, Cursor, ChatGPT — and the Ounie AI Team, which connects with the key in the URL where headers can't be set.

https://research.ounie.com/api/mcp
# header-less clients:
…/api/mcp?api_key=rsd_live_…
x402

No account, no key — pay in USDC on Base. $0.18 quick, $0.42 standard, $0.90 deep.

POST /api/x402/research → 402 (terms)
retry with X-Payment: <signed>
→ 202 { brief_id, poll_url }   # polling is free
Master key

One ounie_live_ key works across the whole Ounie app fleet — including this one.

Authorization: Bearer ounie_live_…
# works across every Ounie app
Read the API docsMint a key

What comes back

Every claim carries the indices of the sources it came from, and every index resolves to a page in sources.

{
  "status": "succeeded",
  "brief": {
    "keyFindings": [
      { "text": "…", "sources": [1, 3] }
    ],
    "answer": [ { "text": "…", "sources": [2] } ],
    "sources": [
      { "index": 1,
        "title": "…",
        "url": "https://…",
        "publishedAt": "2026-07-02T00:00:00.000Z",
        "quote": "the sentence this rests on" }
    ],
    "confidence": 0.78,
    "openQuestions": ["…"],
    "stats": {
      "queriesRun": 3,
      "sourcesConsulted": 14,
      "sourcesCited": 9,
      "claimsDropped": 2
    }
  },
  "markdown": "# …"
}

Credits are reserved before the run and refused with a 402 when the wallet is short, so an agent can never overdraw. A brief with nothing traceable to a source is refunded in full — never settled.

Quick · 1 search
15 cr · $0.18 keyless
Standard · 3 searches
35 cr · $0.42 keyless
Deep · 6 searches
75 cr · $0.90 keyless
Credit math

One wallet. One price per brief. No subscription.

A credit is one cent. Briefs draw the same Ounie wallet every other Ounie app draws, so there is nothing to subscribe to here and nothing to cancel.

  • The price is reserved when you submit and settled when the brief lands.
  • Reading a finished brief, exporting it, and polling a running one are free forever — you already paid for the work.
  • A brief with nothing traceable to a source is refunded in full. We would rather return the credits than hand over an answer you cannot check.
See the full pricingTop up credits
Price per brief by depth, in Ounie credits and US dollars
DepthSources readCreditsCost
Quickup to 615$0.15
Standardup to 1435$0.35
Deepup to 2475$0.75
Ten standard briefs cost $3.50 — about what one hour of the reading they replace is worth. Agents without an Ounie account can pay per call in USDC instead; see the x402 rail.

Questions

Everything about what a brief costs, what it contains, and what happens when the web has no answer.

What stops it inventing a citation?
The model is asked for claims with source numbers attached, and then every number is re-checked against the pages that were actually fetched. A number outside that set is discarded; a claim left with no surviving number is discarded too. There is no path where an unsupported sentence acquires a citation on the way out — the count of what was dropped is returned with the brief so you can see the guard working.
What happens when there is nothing to find?
You are refunded in full and the brief is marked thin, which is deliberately a different outcome from failed. "We searched and the web does not answer this" and "the run broke" are different conversations, and conflating them would hide a real answer behind an apology.
What is the difference between the three depths?
Real work, not prose length. Quick runs 1 search and reads 1,200 characters of each page. Standard breaks the question into 3 angles and reads 3,000. Deep runs 6 angles, reads 6,000 characters per source, and then runs a second pass aimed at the gaps the first pass admitted to.
Can I point it at particular sites — or away from them?
Yes. Pass domains as a list; prefix one with a minus to exclude it. A filter that matches nothing produces a thin brief and a refund rather than a confident answer built on whatever was left.
How fresh are the sources?
Set recencyDays and only pages published inside that window are used. Leave it off and the search is unbounded, which is usually right for questions about how something works and usually wrong for questions about what just happened.
What does it cost?
15, 35 or 75 credits per brief depending on depth. A credit is one cent, drawn from the same Ounie wallet every Ounie app uses. Reading, exporting and polling a brief you already paid for are free.
Can an agent use this without an Ounie account?
Yes — the x402 rail takes USDC on Base with no key and no account. Because an on-chain payment cannot be refunded, that rail runs a real search before it settles: a question with no readable sources is refused, and the agent keeps its money.
Is the answer stored?
The brief and its sources are stored against your account so you can re-read and export them for free. Delete one and it is gone. Briefs paid for over x402 belong to no account and are reachable only by the id returned at purchase.

Ask it something you would otherwise have to check yourself.

A standard brief costs 35credits — about thirty-five cents — and comes back with its sources attached.

Start with your Ounie accountSee what people ask it