Sicut
Early access · in development

Deep research that runs somewhere else, on inference you choose.

When your agent researches something properly, it fans out subagents inside your session — frontier tokens for every page it reads, tens of minutes of waiting, and the whole thing dies if you disconnect. Sicut takes that job off your session and runs it on whichever inference you point it at. Ours, or your own account, at your own rate.

4.6×faster than a fan-out
2-4×cheaper, measured
0context tokens spent reading
Yoursinference, if you want

Head-to-head below, against a real Claude fan-out on the same question.

Sicut enim maius est illuminare quam lucere solum, ita maius est contemplata aliis tradere quam solum contemplari.

For as it is greater to illuminate than merely to shine, so it is greater to pass on to others the things contemplated than merely to contemplate.

— Thomas Aquinas, Summa Theologiae
The whole point

Nobody else will run your research on your own inference.

Every hosted research API bills you for their model. Sicut treats inference as a setting. Point it at Cerebras, OpenAI, Anthropic, OpenRouter, or a model running on your own hardware — it is a base URL and a key, and the bill goes to your provider at your rate.

[ 01 ]

Your inference, your bill

Already have Cerebras credits, an enterprise OpenAI contract, or a box with a model on it? Use them. Sicut works out which model serves which stage from whatever you connect.

[ 02 ]

It outlives the session

Every step is journalled. Close the laptop, drop the connection, come back tomorrow — the job kept running and the stream replays from where you left it. A killed job resumes without re-paying for work already done.

[ 03 ]

It does not eat your context

Reading, deduplication and extraction happen server-side on cheap fast models. Your agent gets back a distilled, cited answer — not ninety fetched pages sitting in the window, re-read on every turn.

[ 04 ]

Or don't set anything up

If you would rather not think about it, use the inference we operate and pay per job. The choice is a radio button, not an architecture.

Measured

The same question, both ways.

A Claude subagent fan-out against Sicut, on one real question with a sixteen-party answer. These are recorded runs, not projections.

Agent fan-outSicut
Wall clock832 s182 s
Cost~$1.17-2.93$0.68
Runs after you disconnectnoyes
Runs on your own inferencenoyes
Pages held in your contextall of themnone
Quotes verified against the sourcenone406 of 406

The fan-out's cost is a range because it depends which frontier model is driving. Sicut's is what the run actually billed, inference and search itemised per call. On a bring-your-own plan the inference half of that is your provider's rate, not ours.

Trust

And you can check the answer.

Cheaper and faster is worth nothing if the report is confidently wrong. Every claim Sicut returns is anchored to a verbatim quote, located in the source document by byte offset — body[start:end] == quote holds, or the claim does not ship. When a search backend refuses, the report says a backend refused rather than letting silence read as an absence of evidence.

sic·ut/ˈsiː.kʊt/Latin, adverb & conjunction
just as; exactly as.
from sic — “thus, exactly so”, the editor's [sic] that marks a quotation reproduced verbatim, errors and all — and ut, “as”.
Plans

Two ways to pay, depending on whose inference it runs on.

Pricing is being set now and will be published before access opens. What is already decided is the shape.

Bring your own inference

A subscription. You connect your provider.

Flat monthly, plus search at costPriced at launch

  • Connect Cerebras, OpenAI, Anthropic, OpenRouter, or any OpenAI-compatible endpoint
  • Point it at a model on your own hardware if you would rather
  • You pay your provider directly, at whatever rate you already negotiated
  • Keys encrypted at rest, never rendered back to anyone, including an admin

Metered

Pay per job. Nothing to set up.

Our all-in cost, plus a marginPriced at launch

  • Pick an inference provider from the ones we operate
  • Every job's real cost itemised — inference and search, per call
  • No monthly commitment
  • A server-side spend ceiling on every job, not a warning after the fact
Connect

One command, then your agent drives.

Sicut is a remote MCP server. Your assistant starts a job, polls it, and reads back the finished report without you opening a browser.

one command
claude mcp add --transport http sicut https://sicut.ai/mcp \
  --header "Authorization: Bearer <your token>"

There is also a page written for the assistant rather than for you — point it at that URL and it will work out the rest.

Access

Not open yet.

Sicut is running in production and is being opened a few people at a time. Leave an address and we will send an invite code when a place frees up.

Already have an invite code? Redeem it here.