How to install SillyTavern: a step-by-step guide

What you are actually installing

SillyTavern is a shell: character cards, chat history, lorebooks and very detailed prompt settings. The text is written not by it but by the model you connect to it. Hence the thing to know first: installing it is not enough — it needs a source of replies, either a model on your machine or a cloud API you bring your own key for.

The steps are below. If your system works differently, check the official documentation: it changes along with the program.

Install these first

Node.js 18 or newer. That is what SillyTavern runs on; take the LTS build from nodejs.org. To check it landed:

node --version

Git. You can technically download a zip, but then every update is a manual re-download. With git, updating is one command.

Installation

Open a terminal (on Windows, Terminal or PowerShell) in the folder where you want the program to live, and take the stable branch:

git clone https://github.com/SillyTavern/SillyTavern -b release

Then start it.

Windows: open the new `SillyTavern` folder and double-click `Start.bat`. The first run takes longer than the rest — it fetches dependencies.

macOS and Linux:

cd SillyTavern
./start.sh

When the terminal reports the server is up, open this in a browser:

http://localhost:8000

That is SillyTavern. Until a model is connected, it will not answer.

Connecting a model

In the interface this is the connection panel — the plug icon in the top bar. Two ways to go.

Your own model on your own machine. Install an engine such as KoboldCpp, Ollama or text-generation-webui, download a model, run it — the engine serves a local address that you paste into SillyTavern. Nothing goes online, nothing is billed monthly, but you need a graphics card, and it decides how long every reply takes.

A cloud model on your own key. Get an API key from a provider, pick that provider in the list, paste the key. There is almost nothing to tune and the quality is high immediately, but every reply costs money and the text goes to someone else's server.

Starting with the second is easier: it separates “the install is broken” from “the model is broken”.

Updating

From the program folder:

git pull

then restart. If you edited files inside the folder, git will say so and refuse — keep your changes in the settings files, not in the code.

Nothing answers: four usual causes

No model connected. The most common one by far. SillyTavern writes nothing on its own — check the connection panel shows a live connection, not an empty field.

The local engine is not running. A model is reached at an address, and that address answers only while the second program is running. Close its window and the replies stop.

Port 8000 is taken. If the address opens something else entirely, another program already sits on that port. The port is changed in `config.yaml`, next to the other settings.

You took `staging` instead of `release`. The first is the developers' working branch and it breaks regularly. If you did not use the command above, check which branch you are on.

About opening it to the network

There is a setting that makes SillyTavern listen beyond your own machine. Before turning it on: by default this is an interface with no password and your API key inside it. Exposed directly to the internet, it will eventually be found by someone else. If you want it on your phone, reach your home network over a VPN rather than forwarding a port.

When none of this is needed

Install, engine, model, prompt, lorebook — that is work, and it pays off in freedom: everything is yours and tuned to you. But if what you wanted was to sit down and play rather than assemble a tool, look at games where the rules are already written: our own Banshee runs the session on d100 rolls with wounds, resources and a hero who can actually die, and there is nothing to install. The ready-made worlds have their own page, and how these tools differ in general is covered in our look at the alternatives.