One of these lives on your phone. The other has never left your desk.

Pick Cosmos if you want to be shown things — it is a discovery app, it is very good at being one, and it runs where you actually have idle minutes. Pick Coii Ref if the pictures you need are already on a drive in your house and the problem is that there are nineteen thousand of them. Neither one is a bad version of the other; they are answers to opposite questions, and the useful part of this page is where the seam between them sits.

Both scorecards, before the argument

Checked against their own listings on 3 September 2026.

The interesting thing about that table is how few rows are contested. Four of them are ours and four are clearly theirs, and the rest are simply about two different places to keep things.

The facts, with where they came from

Cosmos describes itself as a space for inspiration, and the three capabilities it puts forward are worth naming precisely because we have none of them: search by colour, search by visual similarity, and an AI-content control that lets you show, blur or hide generated images. It also makes a point of surfacing the artist, the source and the story behind an image, which in this category is a genuinely uncommon position and not a small one — most of the internet's reference pipeline strips attribution on the first hop.

Their App Store listing is published by Cosmos Entity, INC. and names iPhone, iPad and Apple Vision, requiring iOS 18, iPadOS 18 or visionOS 2 respectively. The in-app subscriptions listed there are $8.00 monthly and $72.00 yearly. The listing was showing version 2.55.0 when this page was written, updated within the past week — which tells you the thing worth knowing about any app you are about to put a collection into.

Coii Ref: $19 USD once, three Macs, macOS 11 Big Sur or later, native on Apple Silicon and Intel, thirty-day trial with every feature and no card and no account.

Where Cosmos is better, specifically

Colour, and finding an image by another image. Two searches we do not have and are not going to. Coii Ref reads colour into the index — that is a fact about the file, alongside dimensions and EXIF — but there is no search-by-colour filter, and saying otherwise would be a lie a reader catches in ninety seconds. If "everything in this bruised green" is the query you want to type, they can answer it and we cannot.

It is with you. Reference-hunting is a queue-and-sofa activity for most people. An app on the phone catches the thing you see at 11pm; a Mac application on a desk in another room does not.

Attribution. A collection that remembers where each image came from is worth more than one that does not, especially if you ever publish work built from it. Coii Ref knows what the file's own metadata knows and nothing more — if the source was lost before the file reached your disk, we cannot recover it.

An opinion about generated images. Being able to hide them, or blur them, or see them clearly labelled, is a real editorial control over a feed. We have no view on it at all, because we only ever show you files you already chose to keep.

Discovery, full stop. Every result Coii Ref returns is something you saved. That is the design, and it is also the ceiling.

On metadata, the honest position: Coii Ref reads what a file already carries — dimensions, EXIF, colour — and makes those searchable. It does not read a source back out of the internet, and it does not invent one. If an image arrived on your disk stripped of everything, as most images that pass through a browser do, then it is stripped here too. The practical answer is the boring one that has always worked: when a reference matters, save the link beside it in the notes field on the file, and it will still be there in four years when somebody asks where the picture came from.

Where the local side is the whole argument

Everything above is about images you have not got yet. This is about the ones you have.

Point Coii Ref at a folder tree and it indexes what is in there without moving anything. No import step, no library format, no second copy, no renaming. The database, the thumbnails and the video proxies are written inside the folder you pointed at, which means the backup you already run covers them and moving the folder moves the workspace:

Then one line of search reaches all of it. AND across keys, OR inside a comma list, no parentheses, and anything the parser does not recognise falls through to free text rather than erroring:

Saved views keep the text of the query rather than a list of files, and re-parse it on open, so renaming a tag quietly updates every view that used it.

And the boards are files. A canvas saves as a .coii-canvas in a folder you chose — pretty-printed JSON that opens in a text editor, diffs in git, and syncs with whatever you already sync with. There is no export because there was no import. Of the three arguments this site makes, that is the only one you can check yourself in ten seconds, which is why it carries the most weight.

What happens at forty thousand files

The reason this is a different product rather than a worse feed is scale of a specific kind: not how many images exist in the world, but how many exist on your disk with no filing system attached.

At two thousand images, folders work and you should not buy anything. At forty, a folder tree is only useful for the material you filed carefully on the day it arrived, which is perhaps one image in six. The rest is real work you paid attention to once and can no longer reach: a client's send from March, a set of scans on the drive you used before this one, the frame grabs from a film you watched twice.

An index changes what a question can be made of. Instead of remembering which folder something is in, you describe the file: how big it is, what kind it is, roughly when it arrived, whether you rated it, whether you wrote anything on it. Those are facts the disk already knows, and none of them required you to tag anything.

type:image width:>=3500 added:>2025-09 no:tags     big, recent, never filed
in:Sends rating:>=4                                 the good ones, one tree
type:video added:>2026-01 has:notes                 clips you wrote on
ext:png tag:any:ui,type -is:favourite               either tag, not starred

Tagging then becomes a small, optional layer over the top: you only need tags for the things the file cannot tell you itself. Mood, project, client, the reason you kept it. Everything else — size, format, date, rating, folder — is already searchable on day one, before you have typed a single tag.

A board you can read without the app

The canvas deserves more than the sentence it got above, because it is the part of this comparison that survives longest.

When you save that board it becomes Bruised greens — study.coii-canvas in a folder you picked. Open it in a text editor and it is pretty-printed JSON: the files it refers to, where each sits on the canvas, how big it is. That has three consequences people notice in order.

The first is backup. The board is inside the tree you already back up, so it is already backed up, and it was backed up the moment you saved it rather than the next time you remembered to export something.

The second is history. A canvas that is text diffs like text. Put the folder in git, or in any versioned sync, and you can see what an idea looked like three weeks ago and go back to it. No product feature was needed for that; it falls out of the format.

The third is the long one. If this application vanished tomorrow, the folder would be exactly what it is today, minus a database and some thumbnails nobody would miss, and the boards would still be readable files. That is not a promise about our longevity — nobody can make you an honest one of those. It is a statement about the shape your work is in, which is the only half of the question anyone can actually check.

The seam, which is the actual recommendation

Most people who land on this page do not need to choose. They need to stop expecting one tool to cover both ends.

Discovery happens on the phone. Cosmos, Savee, Pinterest, Are.na — whichever feed suits your eye. That is where a reference is found.

Work happens on the Mac, from files. The moment an image stops being interesting and starts being something you will draw from, it should be a file in a folder whose name means something to you. Once it is, it joins nineteen thousand others that a query line can reach in one breath — including the ones you saved in 2022 through a service you have since cancelled, which is the part of a collection no feed has ever held.

That habit also answers the question people are usually circling when they compare a subscription app with a one-time one. Not will this company be around, which nobody can answer honestly, but what shape is my work in if it is not. A folder of files and a JSON board is an answer you can verify on any Tuesday.

Video, and why it is not a footnote here

A lot of reference practice is motion now, and most tools treat a clip as a second-class object that lives in a different folder and opens in a different application. Cosmos is an image product; that is a reasonable scope for a discovery feed and not a criticism.

On the local side, video goes through the same index pass as stills. It carries the same tags and the same ratings, it answers the same query line — type:video ext:mov rating:>=4 is an ordinary sentence rather than a separate mode — and each clip gets a proxy written at canvas resolution so a board of footage scrolls like a board of photographs instead of stuttering while several source files decode at once. If half of what you study moves, that is one library rather than two.

What Coii Ref does not have

Said flatly, so nothing in the trial surprises you. No feed, no recommendations and no other people. No accounts, no cloud, no sync, no sharing links, no collaboration, no mobile app — nothing on an iPhone, an iPad or a Vision headset. No browser extension, so saving from a web page stays whatever it already is for you. No search by colour and no search by visual similarity. No search by meaning and no automatic tagging: the query line is literal. No PDFs, documents or 3D formats — images and video. And macOS only, version 11 or later.

What it has instead is narrow and specific: everything already on your disk, searchable without tagging it first; video treated exactly like stills, with proxies at canvas resolution so a board of clips scrolls like a board of photographs; and a network list two items long — a licence check when you press Activate, and one request at launch read only for its Date header so a wound-back clock cannot mint trial days. No analytics, no crash reporting, no update check.

What switching would actually involve

Again: this is not a migration, because there is nothing structurally equivalent to migrate. But there is a sequence that gets a real library standing up in an afternoon.

1. Decide where the tree lives. One parent folder, on the internal disk or an external drive — either is fine, and an unplugged drive reconciles rather than starting over when it comes back.
2. Move the good stuff out of Downloads. This is the only tedious step and it is worth doing once. Folder names should mean something a year from now.
3. Run one index pass over the parent. It works through in the background; nothing is copied and nothing is renamed.
4. Do not tag anything yet. Search what is there first — type:image width:>=2400 added:>2025-06 will tell you more about your own collection in thirty seconds than an afternoon of filing would.
5. Tag only what search cannot already find. Mood, project, client — the things that are not facts about the file.

After that the watcher keeps the index true from filesystem events, so a rename in Finder stays the same item with the same rating and notes, and a folder of new work from a shoot appears without you asking.

Who should pick which

Cosmos if you are short of input rather than short of order: you want to be shown work, you want colour and similarity as search terms, you care about attribution, you want a view on generated images, and you want it on the device you actually hold. $8 a month, or $72 for the year.

Coii Ref if you are drowning in your own archive: the files are already here, the folders are a mess, video needs to sit beside the stills, and you want the boards to be files you own rather than records in a service. $19 once, thirty-day trial, no account.

Both, in the order described above, which is what most working practices actually look like.

Neither, if you have two thousand images and Finder is coping. Buy something when you cannot find a picture you know you saved — not before.

If a desktop application is what you are really after, two neighbours are worth reading: refern is more capable than we are on nearly every axis at $35 once, including an on-device tagger, a search that finds untagged images by meaning, and Windows and Linux builds we do not have; and Apple Photos is already on the Mac and worth understanding properly before you replace it.

Otherwise: the library has the other comparisons and the practical guides, and the trial is thirty days with every feature, no card and no account.