Quickstart

There is nothing to install. chatty-lab runs at chatty-lab.com — you sign in and you are in a workspace.

What follows is the first ten minutes: signing in, giving the platform a key so a model will answer, running a chat, and finding the Studio.

1. Sign in

Authenticate with Google, with email, or with a magic link. The first time, a personal workspace is created for you. Everything you make lands in whichever workspace is active, and the switcher in the top bar says which that is.

A workspace is the unit of everything

Chats, graphs, datasets, experiments, keys, secrets and settings all belong to a workspace — not to you. Switching workspaces changes what you can see and what you are spending. See Workspaces & Teams.

2. Give it a key

A new workspace has no models. That is deliberate: the platform holds no credentials of its own, so what you can reach is what you have configured.

Go to Settings → Providers & API keys and add one. Any of them will do to start — OpenAI, Anthropic, Mistral, Gemini, DeepSeek, GLM, Kimi, Novita, HuggingFace or NVIDIA Build. If you would rather run models on your own machine, connect an Ollama backend instead and skip the key entirely.

The picker tells you what a key actually reaches

Having a key is not the same as being able to call every model behind it. The platform sweeps them and marks in the picker the ones your key does not reach — a model your account has not been granted, or one your Google project has not enabled. See Providers & API keys.

3. Run a chat

Open Chats and start a session. Add a chat to it, pick what it should run on, and send a message; the answer streams back as it is produced.

The thing worth knowing early: a chat does not pick a model, it picks a graph. The model comes from what that graph exposes. That sounds like an extra step and it is the reason the rest works — a chat and an experiment run the same thing, so anything you can say in a chat you can later run over a whole dataset.

Comparing two models is having two chats open

A session holds up to four chats side by side, each on its own graph. Put the same question to all of them at once — that is what the multi-chat layout is for. See Multi-model chat.

4. Find the Studio

Studio is where the graphs live. Its catalog is the spine on the left: every graph, agent, dataset, annotation job and automation the workspace has.

Picking an item shows it. Opening it is a separate gesture — opening a graph joins its collaborative editing session, which is a more expensive thing to do than looking.

If the workspace is empty, Load an example puts a working graph in front of you, which is a faster way to see what a graph is than building one.

Where to go next

Build a graph Add an Input node, wire the catalog, expose parameters, run it and read the trace.

Bind a graph to a chat Make a graph run on every turn of a conversation.

Schedules & doorbells Make a graph run on a clock, or when something outside rings it.