ClawBot Crew · Research Agent

Reading is cheap.Remembering it correctly is not.

Most research tools hand you a confident summary and let you find out later that half of it came from one blog post. The Research Agent works the other way round: it reads what you send, tries to confirm each claim somewhere else, and files only what holds up. What it could not confirm stays visibly unconfirmed instead of quietly becoming part of what you know.

Your wiki belongs to you and grows with every source you share. Nothing is committed to it on a single source alone.

What it actually does

Four behaviours, all of them consequences of one rule: a claim needs two independent sources before it counts as known.

The rule

Two independent sources, or it does not get filed

Every claim it wants to save is cross-referenced against at least two independent sources. Two write-ups of the same press release are not two sources, and it treats them as one.

Intake

It reads what you actually read

Paste a link, drop a PDF, hand it a paper or a podcast. It works through the whole thing rather than the first few paragraphs, and pulls out the claims worth checking.

Recall

Ask your wiki, not the internet

Once something is filed, answering from it takes seconds and costs no re-reading. Your third question about a topic is faster and cheaper than your first, which is the opposite of how a search box behaves.

Honesty

Unconfirmed claims pend, visibly

A single-source claim goes to a pending list with the source attached, rather than into the wiki. You can look at that list. Your knowledge base has no anonymous facts in it.

How a source becomes knowledge

The same four steps run whether you handed it one link or twenty.

01

You hand it a source

A URL, a PDF, a paper, a podcast episode, or just a question broad enough that it should go and find the sources itself.

You
02

It reads and extracts the claims

It works through the material and separates what is actually being asserted from the framing around it. This is the step that decides what gets checked, so it errs towards pulling out more claims rather than fewer.

Agent
03

It looks for a second, independent source

Each claim gets checked against the wiki you already have, your memory, and the open web. The bar is independence: a second article that traces back to the same origin as the first does not clear it.

Claims that clear the bar are written into the wiki with their sources attached. Claims that do not are set aside as pending, with a note on what was missing.

Agent
04

You ask, it answers from what it verified

Answers come from your own wiki plus memory first, and cite what they rest on. When it has to go outside your wiki to answer, it says so rather than blurring the two.

Agent

Try saying

Plain language. There is no query syntax to learn.

What it will not do

The failure modes worth knowing about before you rely on it.

It will not invent a second source

When a claim only ever appears in one place, it stays pending. That is the intended behaviour, and it means your wiki will sometimes be smaller than you expected.

It cannot read what it cannot reach

Paywalled articles, login-gated PDFs and private documents are only usable if you supply the text yourself. It will tell you it could not fetch something rather than summarising the abstract and calling it done.

Verified is not the same as true

Two independent sources agreeing is a much higher bar than one, and still not proof. For anything consequential — medical, legal, financial — the wiki is a starting point and a citation trail, not a professional opinion.

What you need

What it costs

One subscription per crew agent, with its own monthly credit allowance.

Research Agent
$29
per month

Includes 1,500 credits per month for reading, verification and recall. Free trial available — start it from the app and the credits are on us.

Base plan
$9.99
per month, Starter

Crew agents are hired on top of a paid ClawBot plan. One base plan covers every crew agent you hire.

Questions people actually ask

How is this different from asking a chatbot to research something?

A chatbot answers and forgets. Everything it told you last month is gone, and the next answer is built from scratch with no record of what you already established.

This one accumulates. The verification work you paid for in January is still there in June, with its sources, and answering from it is nearly free.

What counts as two independent sources?

Two accounts that do not trace back to the same origin. Five outlets rewriting one company press release is one source wearing five hats, and it is treated that way.

Can I see what it decided not to file?

Yes — ask it what is pending. Single-source claims are kept with their source and the reason they did not clear the bar, so you can go and confirm one yourself if it matters to you.

Is my wiki private?

It belongs to your account and is not shared with other users. Crew agents you own can hand work to each other when you tag them in a chat, which is the only way anything crosses between them.

The rest of the crew

Each crew agent is its own subscription, with its own memory and its own credits. They can hand work to each other when you tag them in chat.

Start with one article

Send it something you were going to read anyway, and look at what it files versus what it holds back. That one comparison tells you most of what you need to know.