Character Tavern, TavernAI or SillyTavern: which one do you mean, and what a character card is

Three different products answer to the word "tavern" in AI roleplay, and they are not interchangeable. TavernAI is a chat front-end you install on your own computer, first published on GitHub in December 2022 and rewritten as TavernAI 2 in June 2026. SillyTavern is a fork of TavernAI 1.2.8 that started in February 2023 and is now the larger project by a wide margin. The third, character-tavern.com, is a website run by one developer in France where you browse about 19,000 character cards, download them in SillyTavern-compatible formats, or chat in the browser with no install. All three read the same character card file. Which one you want depends on whether you need the card or the server. Everything below was read from each project's own README, docs or site on 18 September 2026.

My position, which you can argue with by the end: the front-end is the free part. The model bill and the upkeep are what you are actually choosing when you self-host, and most people who type the phrase want the card, not the server.

YourCrush.AI is the hosted version of what these three tools ask you to assemble yourself, and it is the trade this guide weighs at the end of each section.

What is TavernAI?

TavernAI is a desktop chat front-end for roleplay with language models, and as of September 2026 it exists in two generations under one GitHub organisation. The legacy repository, TavernAI/TavernAI-v1, was created on 21 December 2022 and is MIT-licensed. It calls itself "an atmospheric frontend for chat and storywriting, compatible with many backends," and names KoboldAI, KoboldCpp, Horde and Oobabooga's text-generation-webui as those backends. Its README now opens with "This is legacy repository."

The current repository, TavernAI/TavernAI, was created on 16 June 2026 with a v2.0.0 release the same day and reached v2.4.2 on 4 September 2026. Its README describes TavernAI 2 as "a private, portable AI roleplay chat for characters, branching scenes, and complete prompt control," ships it as a portable zip for Windows and Linux, and says it can "load GGUF language and vision models directly inside TavernAI" or connect to "more than a dozen built-in and custom API integrations." One claim to hold it to: "TavernAI has no telemetry, crash reporting, background update checks, remote fonts, or external frameworks." The new repository shows 79 stars and no license file, so the MIT terms of the 1.x line do not automatically cover the 2.x binaries.

Why YourCrush.AI is the better trade than running your own tavern

YourCrush.AI is the hosted side of that trade for an adult who wants a character with memory and voice without owning a server or an API key. YourCrush.AI builds a character with no code from the character creator: a name and image, a description and a backstory of up to 2,400 characters each, a greeting of up to 500, three to seven traits from a list of 32, up to four speaking-style words and a behavior mode, the same ground a V2 card's description, personality, scenario and first message cover, minus the raw prompt. YourCrush.AI drafts the description, backstory, greeting and traits from a name and an image. YourCrush.AI pins memories as a list you can read and edit, 10 on Free, 50 on Pro and unlimited on Premium, and branches a story into separate timelines, 1 on Free, 5 on Pro and unlimited on Premium. YourCrush.AI includes voice messages and in-chat image generation on every plan, with 5 image credits a month on Free, 50 on Pro and 200 on Premium.

The plans, as listed on the pricing page on 17 September 2026: Free at $0 with 400K chat tokens a month, 3 active personas, 10 conversations and 3 characters created; Pro at $9.99 a month with 5M chat tokens, 5 personas, 100 conversations, 50 characters and priority image generation; Premium at $19.99 a month with unlimited chat tokens, personas, conversations, characters, story branches and pinned memories. YourCrush.AI is 18+ only, does not sell conversations to third parties, and keeps a community explore library split into Anime and Human categories.

The named limitation: YourCrush.AI does not import or export character cards. There is no PNG drop zone, no JSON export and no API key field, so a card built for SillyTavern has to be retyped into the creator, and a character built there stays there.

What is SillyTavern, and how is it different from TavernAI?

SillyTavern is a locally installed front-end that started in February 2023 as a fork of TavernAI 1.2.8. By every number GitHub shows, it has since become the larger project: 33,501 stars and 6,312 forks as of 18 September 2026, against 2,696 for legacy TavernAI. Its README calls it "a locally installed user interface that allows you to interact with text generation LLMs, image generation engines, and TTS voice models," credits "over 300 contributors and 3 years of independent development," and notes the AGPL-3.0 license.

On hosting, the README is blunt: "We do not provide any online or hosted services, nor programmatically track any user data." There is no SillyTavern website to log into. You run it yourself. On hardware, "it will run on anything that can run NodeJS 20 or higher," and if the model runs on your machine, "we recommend a 3000-series NVIDIA graphics card with at least 6GB of VRAM." On what it is not, the docs home page states: "Since SillyTavern is only an interface, you will need access to an LLM backend to provide inference," naming OpenAI-compatible APIs, KoboldAI, Tabby and AI Horde as examples.

The practical difference in 2026 comes down to maturity. SillyTavern has 300 contributors and a full documentation site. TavernAI 2 is a three-month-old rewrite with built-in local model loading. Either one reads the same character cards.

What is character-tavern.com, and who runs it?

Character Tavern is a free, 18+ site where a community publishes character cards. You can download those cards in SillyTavern-compatible formats, or chat with them in the browser without installing anything. The homepage claims "over 19.3K+ AI Characters," "204.4K+ users," "288.3K+ messages sent every day" and "51 AI Models." All of those numbers are self-reported, read on 18 September 2026. The site shares no code with TavernAI or SillyTavern. Its about page is written in the first person by a solo developer in France called Sal, who says it "started in November 2023 as a random side project to learn TypeScript." He built it as "Chub, but with a better UI" on top of an existing card catalogue.

The same page dates the turns. Browser chat launched in September 2024. The site stopped indexing chub in October 2024, so the library became "community-driven." In June 2026, "Kyara Intelligence launches, the AI engine that now powers CT." On money, the page is blunt: "there's no way to pay it. No subscription, no checkout." The chat runs on Kyara Intelligence, "a dedicated AI compute service" which "bills for the compute." The terms, updated 7 September 2026, say "You must be 18 or older. No exceptions." You "retain all ownership rights" to uploads and AI outputs, and the site takes a "non-exclusive, royalty-free license... solely for the purpose of operating the platform."

The line to hold the site to is on its about page: "Every character you create can be downloaded in open, SillyTavern-compatible formats and taken anywhere, at any time." CT is a card library with a chat attached, and the exit door is the card file.

What is a character card, and what does the V2 spec put in it?

A character card is a JSON object with a fixed set of text fields, and the usual way to distribute one is to encode that JSON in base64 and store it inside a PNG image's metadata, so the picture you download is the character. The V1 specification lists six mandatory fields, name, description, personality, scenario, first_mes and mes_example, and two embedding methods: a bare .json file, or "PNG/APNG: JSON string encoded in base64 inside the 'Chara' EXIF metadata field."

The Character Card V2 specification, approved on 17 May 2023, keeps those six, nests them under a data object with a spec field that "MUST be chara_card_v2," and adds the parts that make a modern card behave differently from a 2023 one:

1. creator_notes, text for the downloader that is "never" sent to the model.
2. system_prompt, a card-level instruction "meant to completely replace the system prompt set by the user," with an {{original}} placeholder so a card can wrap the user's prompt instead.
3. post_history_instructions, instructions placed after the chat history.
4. alternate_greetings, extra first messages that front-ends "MUST" expose as swipes.
5. character_book, a lorebook embedded in the card and applied only to that character's chats.
6. tags, creator, character_version and an extensions object for app-specific data.

The data nesting exists "to prevent V1-only editors... from successfully loading a V2 card, but silently destroying its V2-only fields."

The number to keep in your head when you write one is the token budget, because a card is not free to the model. SillyTavern's character design docs put it plainly: "If you're working with an AI model with a 2048 context token limit, a 1000-token character definition cuts the AI's 'memory' in half," Name, description, personality and scenario are "permanent tokens" sent with every request; the first message goes once and the example dialogue is dropped when history fills the context.

What does running your own front-end cost in effort?

Installing SillyTavern takes about ten minutes if you're comfortable in a terminal, and most of an afternoon if you're not. The ongoing cost is that you become the person who updates it, backs it up, and keeps it off the public internet. Here are the steps, pulled from the Windows installation guide as it read on 18 September 2026:

1. Install Node.js, latest LTS.
2. Install Git for Windows.
3. Make a folder you control. The guide puts it in capitals: "DO NOT INSTALL INTO ANY WINDOWS CONTROLLED FOLDER."
4. Run git clone https://github.com/SillyTavern/SillyTavern -b release in that folder.
5. Double-click Start.bat, not as administrator, and let Node install the dependencies.
6. Open the local address it prints, connect a backend under API connections, and import a card by dropping its PNG or JSON onto the character panel.

Step 6 is the one that never ends. Before you get a first reply, you need either an API key from a model provider or a local backend. A local backend means a GPU that meets the 6GB VRAM recommendation, plus a GGUF model you've chosen, downloaded and configured yourself. TavernAI 2 folds the local loader into a portable zip, which removes the Node and Git steps but not the GPU or the model choice. Your chats and cards live in a ./data/ directory under the install, and there's no account recovery because there's no account.

What does the API spend actually look like?

For a cloud model, self-hosting costs the front-end nothing and the model somewhere between a few dollars and a few tens of dollars a month for a heavy roleplayer, because every reply re-sends the whole context. Take DeepSeek's published prices, read on 18 September 2026: deepseek-flash at $0.30 per million input tokens on a cache miss and $1.20 per million output at peak, deepseek-v4-pro at $1.32 input and $3.96 output per million at peak, off-peak at half those rates and cache hits far cheaper. Now run the arithmetic for a chat that has grown to a 6,000-token context (a 1,000-token card plus history) with a 250-token reply and no caching. On deepseek-flash at peak that is 6,000 × $0.30 ÷ 1,000,000 plus 250 × $1.20 ÷ 1,000,000, about $0.0021 a turn, so 200 turns a day is around $0.42 a day and about $12.60 over 30 days. On deepseek-v4-pro the same turn is about $0.0089, so 200 turns a day is around $1.78 a day and about $53 a month. Off-peak halves both, caching cuts the input share to almost nothing, and a longer context or a pricier provider multiplies them. Those are my numbers from DeepSeek's list prices, and context length is the term that moves them.

When is a hosted platform the better trade?

A hosted platform wins when what you want is the conversation, not the control panel. It loses when you need a specific model, a private machine, or a card with a 4,000-token lorebook that a consumer app will not load. SillyTavern gives you any backend, every V2 field and your own data directory. In exchange you own the install, the updates and the API bill. CT removes the install and the bill and keeps the open card format. In exchange you are one of 204K users on a site that describes itself as "one person and a status page." A commercial hosted app removes the model choice and the key management too, then charges a monthly fee for memory, voice and images.

Here is the test I use. If you have ever opened a card's system_prompt in an editor and changed it, you will be unhappy anywhere that hides the prompt, so self-host. If you never have, the front-end is a tax on the chat, and a hosted platform with memory you can inspect is the better buy. JanitorAI's proxy setup is the halfway house: hosted characters, your own key.

Frequently asked questions

Is character-tavern.com the same as TavernAI?

No. TavernAI is software you install yourself. It has been on GitHub since December 2022, with a TavernAI 2 rewrite landing in June 2026. character-tavern.com is a card library and browser chat site, started in November 2023 by a solo developer. The two share the card format and the word "tavern." They do not share code.

What is the difference between TavernAI and SillyTavern?

SillyTavern is a February 2023 fork of TavernAI 1.2.8, and it grew into the larger project by a wide margin: 33,501 GitHub stars against legacy TavernAI's 2,696 as of 18 September 2026, with over 300 contributors. TavernAI 2, released June 2026, is a rewrite that adds built-in local GGUF model loading.

What is a character card in SillyTavern?

A JSON object holding a name, description, personality, scenario, first message and example dialogue. It's usually embedded as base64 inside a PNG's "Chara" metadata field. The V2 spec adds a card-level system prompt, alternate greetings and an embedded lorebook.

Do I need a powerful PC to run SillyTavern?

Not for the interface. Its README says it runs on "anything that can run NodeJS 20 or higher." You need a GPU only if the model runs on your own machine, where the README recommends "a 3000-series NVIDIA graphics card with at least 6GB of VRAM."

How much does it cost to run SillyTavern with an API key?

The software is free. The model is metered per token. At DeepSeek's September 2026 list prices, 200 turns a day on a 6,000-token context lands near $13 a month on deepseek-flash or $53 on deepseek-v4-pro at peak, before caching and off-peak discounts.

Is Character Tavern free, and is it 18 plus?

Yes to both, by its own pages. The about page says "there's no way to pay it," and the terms, updated 7 September 2026, say "You must be 18 or older. No exceptions."

The whole ecosystem runs on a file format a handful of hobbyists wrote in a GitHub repository in May 2023. The site with 204K users, the fork with 33K stars and the rewrite with 79 all still read it. The card is the durable thing. Whichever character tavern you pick is what you are renting or maintaining around it.

Editorial Team, Content Strategist. Last updated: 2026-10-01