Getting Started with Clarifai

Unreachable since 2026 with no shutdown notice: a model catalogue that let you test AI models in a browser tab and then call the same one from your own code.

Last reviewed 2026-09-19 · for Clarifai

How it works

You sign in, browse a library of AI models (the AI "brain" that does the actual thinking), open one in a browser chat box, type a question, and the reply streams back. When you want that same model inside your own software, the site hands you a ready-made code snippet for the exact model you just tried.

What it is

Clarifai is a catalogue of AI models with a control panel wrapped around it. You search the library, test a model in the browser, and — if you want — call that same model from your own software through an API (a way for programs to talk to each other without a person clicking).

It started in 2013 doing image recognition, and that is still the part it does best. The newer work is about *where* a model runs — Clarifai's servers, a cloud you rent, or your own hardware — rather than about building the models themselves.

One thing to know before you invest time in it. In May 2026 Nebius announced it had hired Clarifai's founder and core engineering team and licensed Clarifai's inference technology. Clarifai's older image models and its government work were explicitly excluded from that deal and stayed with the company. No shutdown notice was ever published, but as of 19 September 2026 the platform is not reachable: api.clarifai.com, docs.clarifai.com and status.clarifai.com have no DNS records, clarifai.com resolves to a registrar parking server with dead HTTPS, and www.clarifai.com points at an ISP address that also fails HTTPS. The company's email still routes through Google Workspace. The newest dated release note before the docs went dark was 12.4, from 7 May 2026.

Before you start

As of 19 September 2026 the API host does not resolve. api.clarifai.com has no DNS record, so nothing on this page — the browser Playground, the code snippet, a Local Runner's cloud address — can currently be reached. Treat what follows as a record of how the platform worked, and check whether the host answers again before you plan around it.

Before the hosts went dark, signup asked for no credit card and the Community plan included a small monthly allowance of operations — one that refilled, not a trial credit, but one shared meter that predictions, training and stored files all drew on. None of that can currently be reached, and whether it returns is unknown.

There is nothing to install to begin. Your first hour is a browser tab. A command-line tool only enters the picture if you decide to try Local Runners, which is the last step on this page and entirely optional.

Clarifai is a developer platform, not a chatbot. There is no assistant that remembers you, no conversation history, no mobile app worth having. If what you want is something to ask questions of, start in Text & Conversational AI and come back here when you want to see how models actually get served.

Security & privacy

Clarifai's documentation is unusually direct about the question everyone asks: it states that it does not use your private data to train its models or any other model on the platform unless you explicitly choose to share your inputs and annotations with the Community. So the thing that decides your privacy here is not what you type — it is one visibility setting, and whether you left it alone.

What that means in practice

Sharing to the Community is a licence, not a preference. Clarifai's terms say that by making a repository public you grant every other user a worldwide licence to use and display that content, and grant Clarifai a perpetual, irrevocable right to it. Read that clause before you publish a model or a dataset, not after.
Clarifai calls its credential a PAT. It is an API key (a password that identifies your app, and that spends your money), and it draws on your monthly operations. Keep it out of screenshots, public code and chat messages, and delete and regenerate it the moment you think someone has seen it.
If you try a Local Runner, check whether it is private. Release 12.4 changed the default to private, but the command still accepts a public flag — and "public" here means a public web address pointing at software on your own computer.

What it's good (and not good) at

Best at

Trying a lot of different models in one place, in a browser, without opening an account with each vendor separately
Image and video recognition
Keeping a model and its data on your own hardware while still reaching it through an ordinary cloud address

Okay at

Being someone's first AI tool. The browser Playground genuinely works without code, but everything surrounding it is built for engineers and reads that way.
Free-tier headroom. The Community plan's small monthly operations allowance, at one request per second, was enough to evaluate the platform and not enough to run anything on — and, as of September 2026, cannot be reached at all.

Avoid

Publishing anything to the Community before reading the licence you grant by doing it. Public content is licensed to every other user and, perpetually, to Clarifai.
Treating the Developer plan's promotional price as the standing rate. The pricing page labels it a promotion, and the plan has no published non-promotional price.

First things to try

Test a model in the browser before you do anything else

Opens a model from the Community library in the Playground — a chat box in the browser — and runs it. No key, no install, no code.

Why this one: Clarifai's free meter counts "operations", and predictions, training and stored files all come out of the same monthly allowance. Testing models in the browser is the cheapest thing you can spend it on; uploading a dataset is the most expensive. Doing it in this order means you find out whether the models suit you while the meter is still full.

Tweak: The Playground shows a generated code snippet for whichever model you are looking at. Ignore it for now — you will want it in step four.

Run a picture through one of the recognition models

Uploads an image and returns labels for what is in it, using a model Clarifai trained itself rather than one it is reselling.

Why this one: Image recognition is what Clarifai has done since 2013, and it is the one part of the company the Nebius agreement explicitly left alone — the announcement named Clarifai's legacy computer vision models among the things not included in the deal. If you want to see this platform at its most settled rather than its most uncertain, look here.

Tweak: Use a photo the model should find hard — bad light, an odd angle, something half hidden. Clean stock images tell you nothing about where it breaks.

Optional: point a Clarifai address at a model on your own computer

Installs Clarifai's command-line tool and runs a single local-runner command, which gives a model running on your laptop a working Clarifai web address.

Why this one: This inverts the usual arrangement. The model and the data stay on your hardware and Clarifai supplies only the address and the authentication in front of them — hosted services like Replicate cannot do this at all, because running everything on their own machines is what they are. The free plan includes one runner with a short runtime limit, which is enough to watch it work; the paid Developer promotion is what removes the time limit.

Tweak: This is the only step that needs a terminal, and nothing else on the page depends on it, so skipping it costs you nothing.

Make one call from your own code

Takes the snippet the Playground generated, drops in a PAT created in your account's Security settings, and runs it once.

Why this one: Clarifai's authentication has a specific trap its own docs had to publish a notice about: to call Clarifai's own models you need a PAT rather than an app-scoped key, and you have to pass the literal values clarifai and main as the user and app IDs instead of your own. Almost every "my Clarifai key doesn't work" is this, and knowing it beforehand saves you the hour.

Tweak: Create the PAT only when you have code to put it in. Everything above this step works without one.

Common pitfalls

The free plan allows one request per second. You will never notice this clicking around in the browser, and you will hit it instantly the first time you call the API in a loop.
Use a PAT, with clarifai and main as the user and app IDs — this changed in 2025 and a lot of older tutorials still show the old way.
Don't confuse Clarifai with Nebius Token Factory. Nebius licensed Clarifai's inference technology in May 2026, so the two now describe overlapping capabilities in similar words, but they are separate products with separate accounts and separate bills.

Where to next

More AI APIs and developer services Running open models on your own machine

Want the full picture first? See the Clarifai overview for pricing, the free tier and related tools.