AI character cards: best sites, the JSON format, and where to use them

An AI character card is a small JSON file that describes a character to a language model: their name, how they talk, the situation you meet them in, and the first thing they say. Most are shared as PNG images with that JSON tucked inside the file's metadata, so the picture you download is the character. Chub AI and Janitor AI host the biggest public collections. The part that catches people out is that a card is only half a product. It has no intelligence of its own, so you still need a model behind it, and on most card sites supplying that model is your job.

This covers what is actually in a card, where the good ones are, which apps will run one without you managing an API key, and one thing worth knowing about who owns a card after you upload it.

What is inside an AI character card

Open a card in a text editor and it is unglamorous. The widely used version 2 format is a flat set of text fields that get assembled into a prompt every time you send a message. The ones that matter:

• name and description: who they are. The description does most of the work and is usually the longest field.
• personality: a short trait summary, often a comma separated list.
• scenario: the situation the conversation starts in.
• first_mes: the opening message, which sets tone more than any other single field.
• mes_example: sample exchanges showing how they speak. Cards that feel distinctive almost always have good examples here.
• system_prompt and post_history_instructions: instructions to the model itself, applied before and after the conversation.
• alternate_greetings: other openings you can reroll to.
• character_book: an attached lorebook, which injects extra context only when certain keywords appear.

The PNG trick is why cards spread so easily. The JSON is base64 encoded and stored in a text chunk inside the image, so a card is a normal picture that any site can display and any card aware app can read. That is also why a card you saved from a forum thread still works years later: nothing is hosted, the whole character travels in the file.

It also explains why the same card behaves very differently in two apps. All a card does is contribute text to a prompt. How much of that text survives, how it is formatted, and how much conversation history sits alongside it are decisions the app and the model make. A card is a script, not a performance.

Where to find AI character cards

Four places account for almost everything people actually use.

If you only want to browse and read, any of them will do. If you intend to collect cards and move them between apps, the first three are the real options, because the fourth treats characters as content inside a walled product rather than files you own.

Do you need an API key to use character cards?

Usually yes, and this is the single biggest surprise for people arriving from a Character.AI style product where everything is included. A card site hosts text files. The model that turns those files into conversation is a separate paid service, and on SillyTavern, Chub without a subscription, and most of Janitor's advanced setups, you are the one who signs up for it, generates a key, pastes it in and watches the bill.

That is why the search suggestions around these tools are full of API errors, empty responses, proxies that stopped working and hunts for the best free model. Those are not faults in the card or usually even in the app. They are the ordinary failure modes of a system you assembled yourself: a key that expired, a balance that ran out, a rate limit, a shared proxy that got shut down.

So the useful way to choose is not by which site has the best cards. It is by how much plumbing you are willing to own.

The last row is a genuinely different product rather than a better version of the same one, and it is worth naming the trade plainly. You give up the archive and the ability to swap models. You get billing by card, persistent memory that works without configuration, and a companion built in a designer instead of downloaded. We compare the two approaches in detail on our Chub AI alternative page, including why the crypto only checkout sends so many people looking.

How to write a character card that is actually good

Most weak cards fail the same way: they describe a character instead of demonstrating one. Ten adjectives in the personality field give a model far less to work with than two lines of dialogue in the right voice.

Four things separate cards that hold up over a long conversation from cards that collapse after twenty messages.

• Write example dialogue, and make it specific. The mes_example field is the highest leverage part of the file. Include the character being difficult, not just agreeable, because that is where voice actually shows.
• Give them something they want. A character with a goal will drive a scene. A character with only traits waits for you to drive it.
• Keep the description tight. Every token in the card is a token competing with conversation history for context. Long cards are a common reason a character starts forgetting recent messages.
• Put changing facts in a lorebook. The character_book field only injects entries when their keywords come up, so background detail stays out of the prompt until it is relevant.

The hardest part is usually the blank page rather than the format. If you are stuck on a personality and want thirty rough directions to react to instead of inventing one from nothing, running the premise through a tool that turns a single prompt into dozens of ideas is a faster way in than staring at the description field. Pick the two or three that genuinely interest you and write those properly.

Who owns a character card after you upload it

This gets skipped everywhere and it matters if you write cards you are proud of. Uploading to a public archive is not the same as sharing a file. You are agreeing to a license.

Chub's terms of service, read on 30 August 2026 and marked "Last updated May 15th, 2026", take from anything you upload a "non-exclusive, royalty-free, perpetual, irrevocable worldwide license (with sublicense and assignment rights)". If you set a card public, the terms additionally grant every other user a perpetual, irrevocable license to use, reproduce, distribute and make derivative works of it. Deleting the card later ends the grant going forward, but the terms are explicit that it is not revoked for anything already copied or designated for sublicense.

The same document also states that changes to the terms are "retroactive and apply to any and all users, content and communications, overriding any previously agreed upon terms".

None of this is unusual for a user generated content platform, and Chub at least writes it in readable English. The practical takeaway is small: keep your own copy of every card you write, outside the site. The JSON is portable by design, so there is no reason for the archive's copy to be your only one.

When character cards are the wrong tool

Cards are built for breadth. The appeal is having a thousand characters and starting a new scene whenever you like, and for that they are excellent and nothing else compares.

They are a poor fit for the opposite thing, which is one relationship that accumulates. A card is static text: it says who a character is on day one and it says the same thing on day ninety. Anything the character learns about you lives in conversation history, and history gets truncated. That is the mechanical reason a card based chat feels like it resets, and it is not something a better card fixes. Products built around a companion with persistent memory solve it in the application layer instead, by storing facts about you separately from the transcript.

So the honest split is: if you want variety, collect cards, and pick your app from the table above based on how much setup you will tolerate. If you want one companion who knows you in March what you told them in January, cards are the wrong format and no amount of prompt engineering will change that. Our roundup of the best roleplay AI apps covers the middle ground, and the Character.AI alternatives page is the place to start if you want a large library without running any of the plumbing yourself.

Frequently asked questions

What are AI character cards?

An AI character card is a small file, usually JSON stored inside a PNG image, that describes a character to a language model. It holds their name, description, personality, the opening scenario, their first message and example dialogue. Card aware apps read those fields into a prompt so the model plays that character consistently.

Where can I find the best AI character cards?

Chub AI has the largest public archive and the best search and tagging, Janitor AI has a very large roleplay focused library with chat built in, and the SillyTavern community circulates the most technically elaborate cards through Discords and subreddits. For cards that work with no setup at all, the in app libraries at Character.AI, SpicyChat and PolyBuzz are easier but not exportable.

How do I use a character card without an API key?

Pick an app that includes the model. Character.AI, SpicyChat, PolyBuzz and Janitor's basic tier bundle inference into the product, so a character works the moment you open it. Chub includes models only on its $20 a month plan, which takes cryptocurrency and not cards. SillyTavern always needs a backend you supply.

Can I convert a character card between apps?

Usually yes. The version 2 card format is a shared standard, so a PNG exported from one card aware app generally imports into another, and converters exist for older formats. What does not transfer is the conversation. Chat history, memory and any app specific settings stay behind, so you get the character back but not the relationship.

Are AI character cards free?

The cards are. Every major archive lets you download them without paying, because they are community made text files. The recurring cost is the model that runs them, whether that is an API bill from OpenAI, Anthropic or Google, a subscription to a platform's own models, or the electricity and hardware for running one locally.

Why does my character forget things mid conversation?

Because the card and the recent conversation share one fixed context window. As the chat grows, older messages get pushed out to make room, and a long card description takes space that history would otherwise use. Trimming the card, moving background detail into a lorebook, or moving to an app with real long term memory all help, in that order of effort.