Roleplay clients
No account exists, so there is nothing to attach your chats to.
A base URL and a key for SillyTavern, RisuAI or Agnai. You pay to an address and receive a key; the key is the identity. Prompts and completions are not stored by default — not as a setting we promise to respect, but because the code that would store them does not run.
- Signup
- None
- Prompt retention
- Not stored by default
- Minimum deposit
- $5
- Recovery
- Two saved secrets
No email, no name, no phone. There is no account to create.
The storage code does not run. Logging is off unless you turn it on.
Credits never expire, so a month away costs nothing.
Your key and a permanent dashboard link. Lose both and nobody can help.
What we keep and what we never see
Stated precisely, because a vague privacy claim is worth nothing to the people who read these pages carefully.
Kept
A usage record per request: token counts in, out and cached, the model id, the upstream that served it, the cost, the latency and the status. That is what billing and the activity log are built from, and it is the whole list.
Not kept
The character card, the lore book, the persona, the messages and the replies. They are relayed and forgotten. Logging controls exist in settings and default to off.
One thing this does not cover. Your prompts still transit the upstream provider that serves the request, and each upstream has its own retention policy — we do not speak for theirs. The upstreams that can serve a given model are named on its model page, so you can check before you choose. The full boundary is in /docs/privacy.
Clients · 3 guides
All three take a custom OpenAI-compatible endpoint. The failure that eats the most time in every one of them is a base URL with /chat/completions appended by hand — the client adds that path itself.
Every guide here is a draft. None of these clients has passed our recorded-fixture compatibility suite, so the field names are our current reading rather than something replayed in CI. Run the one-request verification on each guide before trusting a block.
Why roleplay bills strangely
A roleplay turn is a very long input against a very short output, and then you regenerate it.
The card, the lore book, the persona and the whole visible history get re-sent every single turn, while the reply is a couple of paragraphs. That inverts the usual assumption that output tokens dominate a bill, and it changes which model is actually cheapest for you — a model with a low output rate and a high input rate is the wrong way round for this workload. The arithmetic, against current catalog rates, is in what a roleplay message actually costs.
Models worth trying
Open-weight models picked for prose and for holding a persona across a long thread. These are judgements — we have no usage data to rank by, and will not pretend otherwise.
nousresearch/hermes-4-70bInstruction-following candidate that holds a persona across a long thread.sao10k/l3.3-euryale-70bProse-oriented candidate. Single upstream — see the note on the roleplay facet.meta/llama-3.3-70bGeneral-purpose prose candidate.mistral/mistral-small-3.2Smaller-model candidate for frequent regeneration.
The full cut, with context windows and current rates, is at /models/roleplay. Frontier models from the large labs are bring-your-own-key here: their terms prohibit reselling, so our resale inventory is open-weight only.
Paying without an account
Prepaid credits, bought with crypto. Which rails are actually available is reported by the payment backend on the deposit screen rather than promised here.
Save both secrets before you pay. Accountless access rests on two bearer secrets: the API key your client uses, and a permanent dashboard link for the browser. The link can replace a lost or rotated key without us collecting an identity. Lose both and there is no personal recovery — we are telling you this before payment rather than after, because a constraint discovered later reads as a trick.
Deposits only, no withdrawals. Credits buy inference and never convert back to coin. That is a deliberate regulatory choice and it is what keeps the product KYC-free: a withdrawal path would make us a money transmitter, and money transmitters must verify identity. Refunds are service credit, never cash.
What this is not
No retention and no account are privacy properties, not permission. The models we serve carry whatever their own weights and the serving upstream impose, and our acceptable-use policy applies the same way it does everywhere else on the service. We would rather say that plainly here than have you find out from an error.
Next
- Roleplay modelsThe curated cut, with context windows, rates and who picked it.
- Cost per messageWorked arithmetic for a long prefix against a short reply.
- SillyTavern setupThe Custom (OpenAI-compatible) source, field by field.
- No KYC, structurallyWhy there is nothing to verify, and what that costs you.
- Data handlingThe exact usage record, and where the provider boundary sits.
- Personal agentsThe other cluster that re-sends a long prefix on every turn.