Build on your own research
Everything you keep here is reachable from a script, a tool you are writing, or a model you have given a key to. Your notes, your boards, the sources behind each claim, and the archived copy of every page you have ever cited.
How it works
A key acts as you, not as an application
There is no second permission system to learn and no scopes to map onto your own work. A key is exchanged for your own session, so it reads exactly what you read and writes exactly what you may write. A private notebook is absent from the API for the same reason it is absent from somebody else's dashboard, rather than because a filter was remembered.
That is a safety property more than a convenience one. The usual way this goes wrong is an integration that quietly holds more access than the person who installed it. Here there is nothing for it to hold.
What it is for
Three things people actually do with it
Put your reading where your claims are
Pipe anything that finds you a link — a bookmark queue, a feed, a colleague's email — straight onto a board as a note with the source already cited. The page is archived at the moment it arrives, so what it said that day is kept whether or not it survives.
Let a model draft, and sign it yourself
Give an agent a key and let it search, read what a note rests on, and propose sources. Everything it suggests waits for you to accept it, including on your own notes, so a model can be useful without anything it produces becoming a claim you never read.
Take a board somewhere else
Walk a notebook, its sections, its notes and their citations, and render it into whatever your work actually ships in — a brief, a slide, a filing, a site. Each citation comes back with a link to the archived page on this site, so a reader of your output can still check the source.
What you get that you would otherwise build
The archive comes with the citation
Cite a URL through the API and it is captured: the page, its text, and a record of it changing afterwards. You are not standing up a crawler, deciding how often to re-check, handling the half of publishers that refuse a datacenter address, or storing any of it.
Captured pages are read on this site and never handed over, which is what makes keeping them defensible. A citation carries the address, the publisher, whether the page has changed since, and a link to the capture here.
For teams
What a business gets out of it
Boards a team makes belong to the team, so an integration built around them keeps working when the person who wrote it leaves. Keys are issued per person and revoked in one press, and each one can be narrowed to a single board — which is the difference between a key leaked into a repository costing you that board and costing you everything.
A key records when it was last used, including the attempts that failed, because “has anyone touched this in six months” is the question somebody asks while deciding whether to revoke it.
Guarantees
Four things it will never do
Publish anything
Notes and boards made through the API are private, and the field is not read from the request at all. Nothing you build, and no model holding your key, can make your work public by accident.
Decide what you have accepted
A source suggested by a model-held key waits for you, even on your own note, and it does not clear the unsourced flag until a person has looked. That is enforced in the database, not in the client.
Delete across other people's boards
It can delete a note only you hold. One sitting on somebody else's board is refused with a count of the boards holding it, and points at abandoning instead — your name comes off and the note stays where people put it.
Hand over the archive
Captured pages are read here and never downloaded in bulk — by you, by anybody integrating, or by anybody who finds a key.
Work from a web page
There is no cross-origin access, deliberately. A key belongs on a server; allowing one from the browser would let any site a person visits spend theirs.
Agents
Or hand it to a model
A skill file at /ai-skill teaches an agent how this place works, and a key can be marked as model-held when you issue it — after which everything it proposes waits for you, including on your own notes.
That is enforced in the database rather than asked for in a prompt, which is the difference between a safeguard and a hope.
Plans
What it costs
The API is included rather than sold separately. It needs a paid plan, which is how we know there is a person behind the requests, and what moves between tiers is how much of it you may use.
- Researcher
- $7 a month
- 120 a minute500 writes a day250 new pages a month
- Business
- $12 per seat
- 600 a minute5,000 writes a day2,500 new pages a month
Teacher carries Researcher's figures. Every answer tells you how much of the minute is left, so a client never has to find the limit by crossing it, and the other two refusals name the number they were measured against. Student accounts can never hold a key.
What is counted
A page, not a request
Requests are cheap and the minute limit exists to stop one runaway script, not to ration you. The figure worth reading before you build anything is the other one: how many pages a month you introduce to the archive.
That is the one call here that buys an open-ended obligation. A page nobody has cited before is fetched in a real browser when the publisher refuses a plain request, which roughly half of them do; the page, its text and a screenshot are stored; and from then on it is re-checked for as long as the citation exists, on a cadence that gets faster the more it is cited. Nobody can say in advance what one page will cost us over five years, so it is the thing with a number against it.
Which is also why most of what you do is free. Sources are deduped by address, so the thirty-first person to cite an article spends nothing: the archive already holds it and their citation points at the same capture. The allowance is only for pages this platform has never seen.
It is counted against the account rather than the key, so ten keys share one allowance and revoking a key resets nothing. It starts again with your billing month. Writes are counted the same way and reset at midnight UTC.
Start with one request
Issue a key in settings, and the first call is one line. The reference has every endpoint and the four refusals worth knowing about.
curl https://critically.io/api/v1/notes \ -H "Authorization: Bearer YOUR_KEY"