Affiliate disclosure: this page contains affiliate links. We earn a commission when you sign up through our links, at no extra cost to you. Our editorial opinions are never paid for. Adult content 18+.
SillyTavern is not an AI girlfriend app. It’s a program you install on your computer that opens in your browser like any other tab. It has no model of its own, no characters of its own, and it doesn’t charge you a cent: what it does is hand you the most complete control panel that exists for talking to a language model, with the character card you pick and the AI provider you pay for.
That’s the honest definition, and it also explains why half the people who try it walk away within two days. Here’s what you actually need to know before you dive in: which pieces you have to assemble, what it really costs, what it does better than a closed app, and the point where it turns into a second job.
SillyTavern in 30 seconds: what it is and what it isn’t
The official documentation describes it as “a locally installed user interface that allows you to interact with text generation LLMs, image generation engines and TTS voice models.” Translated: it’s the steering wheel, not the engine.
| What SillyTavern is | What SillyTavern isn’t |
|---|---|
| A web interface running on your own machine (NodeJS 20 or higher) | A mobile app you download from a store |
| Free and open-source software, AGPL-3.0 license, no cost | A service with its own monthly subscription |
| A manager for character cards, lorebooks and personas | An AI model trained for roleplay |
| A client that connects to whichever API you choose | An account or hosted content generator |
| A community project with more than 300 contributors | A product with paid official support |
It started as a fork and is still a hobby project
SillyTavern began in February 2023 as a fork of TavernAI 1.2.8 and has run on volunteer contributions ever since. The team maintains two branches: release, updated roughly once a month and the one you should be using, and staging, updated several times a day and liable to break without warning. They don’t take donations for the project, which says a lot about how it’s put together: nobody is selling you anything inside it.
The missing piece: it doesn’t ship with a model
This is the first thing that hits you. You install SillyTavern, open it, and there’s no AI. The interface sits empty because you supply the model, either by plugging in a cloud API or by running a model locally. That one decision is what sets the cost, the quality of the replies, the content filter and the privacy of everything you type.
How to install it, depending on the machine you have
There’s no double-click installer that works for everyone. Four officially documented paths exist, and each one has a clear profile.
| Platform | What you need | Difficulty | When to pick it |
|---|---|---|---|
| Windows | Git, NodeJS 20+, a terminal | Medium | It’s your main PC and you want the full experience |
| Linux / macOS | Git, NodeJS 20+, a terminal | Medium | You already live in the terminal and a git pull doesn’t scare you |
| Android (Termux) | Termux from F-Droid (not the Play Store), git, nodejs-lts | High | You want your characters in your pocket without depending on the cloud |
| Docker | Docker installed and some comfort with containers | Medium-high | You want to isolate the environment, serve it on your local network, or update without messing up your system |
Windows: the short path
You clone the repository on the release branch, install the dependencies and launch the startup script. It prints a local address (port 8000 by default) that you open in your browser. The whole time you’re working against localhost, not somebody else’s server.
Android with Termux: it works, but the hardware shows
The mobile route goes through Termux installed from F-Droid or the GitHub releases — the Play Store version is abandoned — then pkg install git nodejs-lts nano, cloning the repo, and bash start.sh inside the folder. On 32-bit Android there’s a known error (Unsupported platform: android arm LEtime-web) that you fix by installing esbuild separately. The documentation itself recommends editing config.yaml to lighten the memory and disk load (lazyLoadCharacters: true, useDiskCache: false), because a phone just can’t take more.
Docker: the option for people who already know
If you already run Docker, this is the least painful path over the medium term: you update with a new image and don’t drag along odd dependencies. In exchange, you have to understand volumes and ports, and your setup becomes one more file to maintain.
Connecting the model: the four real routes
This is where your monthly bill gets decided. SillyTavern supports dozens of backends; in practice they come down to four decisions.
| Route | Cost | Quality | Privacy | Difficulty |
|---|---|---|---|---|
| AI Horde (built in) | Free, shared queue | Variable, depends on who donates capacity | Low: your text passes through volunteers | Very low |
| Cloud API (OpenAI, Claude, Google AI Studio, Mistral, DeepSeek, OpenRouter) | Pay per token | High | Low to medium: the provider reads the prompt | Low: paste a key |
| Local model (KoboldCpp, llama.cpp, Ollama, TextGeneration WebUI) | Free after the download | Medium, lower with small models | High: nothing leaves your machine | High: 5 to 50 GB per model |
| Specialized provider (NovelAI, Mancer, DreamGen, Pollinations) | Subscription or credits | High for narrative | Medium | Low |
The real cost of the cloud route
The software is free; the tokens aren’t. A top-tier model charges per million tokens, and a long roleplay session burns far more than a single question because every message resends the accumulated context: character, lorebook, history and notes. If you roleplay an hour a day on an expensive model, the bill looks closer to a platform subscription than to a free game. The upside is that you pay exactly for what you use, and you can drop to a cheaper model whenever you like.
The real cost of the local route
Zero in API fees, plenty in hardware. The documentation recommends an NVIDIA 3000-series card with at least 6 GB of VRAM if you’re going to run inference on your own machine. With less it still works, just slowly. And even though local models apply no commercial filter, their conversational quality usually sits below the big cloud models.
What it does well (and why people pick it)
If all of this were nothing but complication, nobody would use it. These are the features where a closed app falls short.
World Info: the lorebook that keeps things consistent
You can create context entries that trigger on keywords: a neighborhood, a side character, a rule of the world. Instead of stuffing everything into the character description and spending context on every single message, the lorebook injects only what’s needed, when it’s needed. It’s the difference between a character who remembers their past and one who improvises it.
Importable character cards
The card format (V1, V2 with the chara chunk inside the PNG, V3 in a .charx container) was born in this ecosystem. SillyTavern reads them, exports them, and lets you move a character between tools without rewriting it. If you’re coming from a closed app, this is the detail you’ll appreciate most.
Group chats, personas and RAG
Group chats put several characters in the same room, talking to each other and to you. Personas tell the model who you are, so it doesn’t treat you as a generic user. And the Data Bank adds documents to the conversation with built-in retrieval (RAG), which is handy when you want the character to consult your own notes.
Extensions: voice, images, translation
The extension ecosystem covers what the base app doesn’t: emotional expressions with sprites, automatic history summaries so you don’t lose memory, translation of the interface and of the chats, image generation with Stable Diffusion, FLUX or DALL·E, text to speech for the replies, and web search to give the model real context. All of it installs from the extensions menu itself.
Where you’ll get stuck: five real friction points
This is the section that rarely shows up in enthusiastic tutorials.
1. The terminal is not optional
If you don’t know what a git clone is, you’re going to struggle. It’s not impossible — there are step-by-step guides — but the mental model of “installing a program” is different here: dependencies, Node versions, startup scripts.
2. Updating can break things
The release branch is stable, but extensions run on their own schedule. A change to the card format or the prompt template can leave a chat giving you strange answers. The community handles it with manual backups of the data folder, which is not what a hosted app does for you.
3. Memory isn’t infinite, it’s configuration
The model only sees what fits in its context window. Everything else is on you: automatic summaries, a well-built lorebook, trimming the history. There’s no magic here, only settings. If you’re coming from an app that “remembers” for you, the learning curve is real.
4. The filter depends on the backend, not the interface
SillyTavern doesn’t moderate: moderation is applied by the model provider. With local models there’s no commercial filter; with big APIs there is. That means behavior changes when you change backends, and that “unfiltered” isn’t a feature of the program but a consequence of what you connect to it.
5. On mobile you pay in battery and patience
Termux works, but it isn’t a native app: you have to keep the process alive, fight Android’s power saving, and accept that the phone gets hot. Fine for short sessions; tiring for daily use.
SillyTavern vs a hosted app: when each one wins
There’s no absolute winner. There are profiles.
| Criterion | SillyTavern | Hosted app (Candy AI, Chai, Character.AI) |
|---|---|---|
| Setup | Manual, 20-60 minutes the first time | Zero: sign up and chat |
| Cost | Free software plus API or hardware | Monthly subscription, free plan available |
| Prompt control | Total: templates, format, parameters | None or very limited |
| Characters | Yours, imported or created | The platform’s catalog |
| Voice and images | Through separate extensions and APIs | Built into the subscription |
| Privacy | High locally, depends on the backend | Medium: the provider hosts the conversation |
| Mobile | Termux, no native app | App or responsive web |
| Support | Community (Discord, GitHub) | Service support |
When SillyTavern wins
When you want fine control over the prompt, portable characters, large lorebooks, group chat and zero dependence on a platform that can change its rules or its business model. If you already pay for an API for other work, the marginal cost is close to nothing and the freedom is total.
When a hosted app wins
When you want to talk to a character tonight, with nothing to install, with voice, images and a predictable bill. Candy AI, for instance, runs in the browser with a limited free plan (around 30 messages a day, no NSFW and no image generation) and a Premium plan from $5.99 a month on annual billing, with 100 monthly tokens included for photos and voice. The token system is broken down in the Candy AI pricing guide: text chat doesn’t consume tokens, image and voice do.
Put another way: SillyTavern gives you the engine and the steering wheel; an app gives you the whole car with a full tank and the fine print of the contract.
Migrating your characters to SillyTavern without losing them
If you already have characters in another tool, the move is simpler than it looks, as long as the platform lets you export.
Step 1: get the card, not a screenshot
You need the real file (a .png with JSON inside, a .json, or a .charx). A screenshot is useless: without the text fields there’s no character, just a pretty picture.
Step 2: check which format you have
If it’s a V2 PNG, almost any reader will open it. If it’s .charx (V3), you need a recent version or a conversion. If it’s a loose JSON, it’s probably V1 and you’ve lost the creator’s notes and the alternate greetings. The guide to AI character cards covers which field each version stores.
Step 3: import it and check the five fields that actually matter
The model doesn’t read all fifteen fields in the card. The ones that drive the conversation are name, description, personality, scenario and first message (plus the dialogue examples). If the character feels flat after importing, it’s almost always because one of those five is empty or truncated.
Step 4: rebuild what the app doesn’t export
Chat history, accumulated relationships and the platform’s internal settings don’t travel. You can fill the gap with a lorebook that summarizes what matters, but don’t expect perfect continuity on day one.
FAQ
Is SillyTavern free?
Yes. The program is free software under the AGPL-3.0 license and has no paid version or subscription. What can cost money is the model: if you connect a cloud API you pay per token, and if you run it locally you pay in hardware and electricity.
Do I need a character card to chat?
Yes, for a persistent conversation. SillyTavern is built around the card concept: it’s the set of instructions that defines how the model behaves. You can type on the welcome screen without picking a character, but that creates an empty card you’ll have to customize afterwards.
Does it work on iPhone or Android without a computer?
On Android, yes, through Termux, installed from F-Droid or GitHub rather than the Play Store. On iPhone there’s no officially documented route: the usual approach is to run it on a computer and open it from the phone’s browser on the same network.
Does SillyTavern remove the content filter?
Not exactly. SillyTavern applies no moderation of its own, but the conversation passes through whatever backend you connect. With a local model there’s no commercial filter; with a cloud API there is, and it depends on that provider’s policies.
How much memory does a character have in SillyTavern?
As much as fits in the context window of the model you’re using. There’s no infinite memory: you decide what goes in (description, lorebook, history summary) and what gets trimmed. A well-built lorebook pays off more than upgrading to a bigger model.
Is it worth it if I just want to chat for a while?
If your case is “open it and chat today”, no. The first install takes 20 to 60 minutes and you still have to sort out the backend. For that, a hosted app with a free plan is more practical; if you also want built-in voice and images, see how the best AI roleplay apps compare before you decide.
Related articles: How to create an AI character that doesn’t forget anything · Chatting with AI characters: what to expect · Chub AI alternatives · Character AI alternatives for roleplay · AI waifu chatbot · Full Candy AI review