help & contact

new here? start here ✦

Make a free account, bring your own AI provider key, and start chatting in a few steps. The getting-started guide has the whole tiny ritual.

This is a brand spankin' new site as of July 2026, and you're bound to find bugs. Whenever you do, email [email protected] or send a message to Zumi.

We're archiving some of our favorite accounts from CharacterHub. If you see your content here, feel free to message us and we'd be glad to move the account over to your email address. (We'll have you add a note on your CharacterHub profile so we know we're handing over the account to the right person.)

moving from another character site? grab a box 📦

Bring your cards and chats, keep the old setup while you settle in, and make a backup whenever you want. The moving-to-CardChan guide has the whole packing list.

ok so what's a "shadowban"?? 👻

Sometimes a card (or, rarely, a whole account) gets quietly hidden from browsing and search. That's a shadowban, and bestie, we need you to hear this: it is not because we hate your card. Your card might be gorgeous. Your card might be your magnum opus. But some of our users are, how do we say this lovingly… delicate flowers 🌸 — and if certain content blooms in their browse page, they will absolutely gather at our firewall with pitchforks, and then nobody gets a nice website.

So to keep CardChan usable for everybody — the spicy connoisseurs AND the faint of heart — sometimes we hide things instead of deleting them. A shadowbanned card is still yours: you can see it, edit it, and download it. It just stops showing up for everyone else.

If one of your cards gets limited, you'll get a message about it (no, we won't tell you which mod did it — don't ask, that's the whole point of the pitchfork thing). Think it was a mistake? It happens! Head to appeals and plead your case — we genuinely do un-hide things when the vibes check out. 💅

ok so what's a "Mythic" ✨??

Every so often a creator's work is so consistently unhinged-in-the-best-way that we hand them a holographic badge and stop rate-limiting their creativity. Mythics have achieved remarkable status as creators — think of it as CardChan sainthood, minus the martyrdom. It's earned by making great stuff, not by moderating anyone else's — Mythics can't touch other people's cards, they just get to make more of their own, faster. Powers:

  • ✦ No rate limit on AI card generation (fork remixes + "ask AI for help")
  • ✦ No rate limit on AI cover image generation
  • ✦ No rate limit on uploads, downloads, or tagging
  • ✦ Bigger batches when bulk import lands

Mythic status shows up as a shimmery ✨ badge on your profile and in the creators directory, plus a Mythics filter in Craft Tier search. It's granted by the site owner and, once given, it's yours for good.

what happens when something gets a safety flag?

Signed-in members can flag a card, collection, lore book, prompt, or creator profile that may be unsafe. That exact version disappears for the person who flagged it and for public visitors while a moderator and the site owner check it separately. The creator gets a message right away.

If both reviews clear it, it stays up and cannot be flagged again unless its safety-relevant content changes. If the final decision rejects it, it stays hidden and the creator gets an exact seven-day deadline to submit a safer edit. If they do not, the creation and its stored files are permanently deleted. That deletion cannot be undone.

The hard lines are graphic NSFL content, sexualized or exploitative depictions of minors, and written depictions of public figures. The flag menu lives on each supported page — no public callouts, no reporter names, just a careful review.

are there NSFW cards here?

Yes, but you have to opt in — nothing spicy shows up until you're 21+ and flip the switch yourself in settings. Tried to view something we gated? Here's the full explainer.

about the AI leaderboard 📊

The AI model tea page rolls up real CardChan runs by provider and model so your BYOK picks have receipts, not vibes. Every number on it is gated so tiny samples can't out you or mislead you — here's what all the thresholds actually mean.

who shows up at all

A provider+model row only appears once it clears a privacy floor: at least 3 different people used it that day, or 25 total attempts in the window. Anything quieter stays hidden and gets counted in the little "stayed private" note under the table, so you know the board isn't the whole story.

stars, keeps, and continues

A model needs at least five ratings before its stars can lead the sort — one gushing five-star vote shouldn't crown anybody. Swipe keep rate needs 20 keep/discard decisions before it means anything. For chat replies, the continued-from rate needs 20 decided replies, and brand-new or untouched replies stay undecided and out of it — that's a separate signal from swipe keeps. Room stars (how a whole group chat rated an image) need 5 ratings from 3 different chat members to show or lead.

images: survival and spice

The delete/replacement rate for generated images needs 20 successful generations to appear, and it's forward-only since launch — an image you never decided on isn't counted as a keep. The NSFW percentage is measured among surviving chat images only; anything you deleted is already gone and doesn't count.

vision: drafts and safety calls

For image-to-card, the accepted-draft rate needs 20 explicit keep/forget decisions, and drafts you never touched stay out of it. For NSFW classification, the quality report covers the human-adjudicated subset only and needs 50 decisions — including at least 10 actually-NSFW and 10 actually-SFW examples — before accuracy, precision, and recall show. How reliably a model answers and how accurately it classifies are two separate receipts; don't read one as the other.

whole-route reliability

Request experience gets attributed to a model only when it was your first-choice pick, not when it caught a failover. Route metrics — first-choice success, failover use, end-to-end time — need 3 daily users or 25 routes to go public, and "collecting" never secretly means 0%.

cost and speed, honestly

Reported cost needs 80% coverage before it can lead the cost sort — a model that only tells us its price one call in ten doesn't get to win on price. Route cost needs 80% fully-covered completed routes for the same reason. The ~p50 and ~p95 timings are approximate estimates from provider-call histograms, and time-to-first-text (TTFT) needs 20 measured streams and 80% coverage before it shows.

Refusal rate looking dramatic on a model you like? Here's how to give the robots an easier prompt so they say yes more often.

Need a hand, or want to report something?