SillyTavern API provider setup
Configure SillyTavern against an OpenAI-compatible endpoint in four fields, then pick an uncensored model for long roleplay sessions.
SillyTavern is the most widely used frontend for character chat and roleplay. GeniusRelay works as its API provider over an OpenAI-compatible endpoint, which means your existing characters can run on 599 models, including 43 uncensored editions, without a proxy in between.
This guide is the short version. The maintained step-by-step with screenshots lives at SillyTavern setup.
Before you start
- A GeniusRelay key from the console. Give it a spend cap; roleplay sessions are token-heavy.
- SillyTavern installed and opened locally.
- One character card. Start with a short card while testing; long cards add input tokens on every turn.
The four settings
In SillyTavern open API Connections and fill:
API type: Chat Completions
Custom endpoint: https://api.geniusrelaygpt.com/v1
API key: YOUR_GENIUSRELAY_KEY
Model: deepseek-v4.1-flash-uncensoredThen click Connect. A green status means the endpoint answered. If the model dropdown is populated, the key is valid.
Which model to pick
Roleplay punishes refusals and forgetfulness at the same time, so the model needs both a high tolerance for odd prompts and a long window.
| Model | Input / 1M | Context | Why |
|---|---|---|---|
| DeepSeek V4.1 Flash Uncensored | $0.429 | 1M | strongest uncensored option |
| DeepSeek V4.1 Flash Heretic | $0.257 | 512K | cheaper, still few refusals |
| GLM 5.3 Flash Uncensored | $0.260 | 1M | bilingual, long context |
| MiMo V2.6 Flash Uncensored | $0.650 | 1M | light and fast |
| Gemma 4 26B Chimerax Heretic | $0.156 | 262K | cheapest entry point |
The roleplay models page lists all 96 roleplay-tagged models with live prices. Sort by input price if budget matters more than the last few percent of quality.
Context settings that save money
Every turn resends the whole history. A 30K-token chat on GLM 5.3 Flash Uncensored costs about $0.008 per turn in input alone; at 500 turns that is $4 before any output. Tune these before a long session:
- Context size: set it to what the card needs, not to the model maximum.
- Summarize older turns: let SillyTavern compress history instead of resending it raw.
- Trim world info: each always-on entry is permanent input tokens.
- Set `max_tokens`: 300–600 is usually enough for a chat reply; 2,000 invites rambling and a bigger bill.
Use the pricing page if you want to model a session length before starting.
Streaming and formatting
Streaming is supported and makes long replies feel immediate. Leave it on unless a reverse proxy buffers events.
For consistent output, keep the response format plain text and avoid forcing JSON in a roleplay prompt. If a model starts emitting markdown headings inside dialogue, add a short system instruction instead of switching models.
Troubleshooting
| Symptom | Likely cause | Fix |
|---|---|---|
| 401 | key copied with whitespace, or revoked | regenerate the key in the console |
| 404 on connect | endpoint missing /v1 | use https://api.geniusrelaygpt.com/v1 |
| 429 | per-key rate or spend cap reached | raise the cap or slow the retry loop |
| Empty model list | connection failed before the key check | reconnect, then refresh |
| Replies cut off | max_tokens too low or context overflow | raise the cap or trim history |
| Repeated refusals | aligned model selected | switch to a model from uncensored |
Full error shapes are in error codes.
Character cards and memory
Long cards raise quality and cost. Two practical compromises:
- Move stable lore into a summarized block at the top instead of a wall of example dialogue.
- Keep the greeting short. The first message is resent every turn.
If a card needs more than 100K tokens of setup, use a 1M-context model and accept the input bill, or split the lore into world-info entries that only trigger on matching keywords.
Is it against the terms?
No. SillyTavern is a client; GeniusRelay only sees normal chat completion requests. Billing is the same as any other client.
FAQ
Does GeniusRelay support streaming?
Yes. Standard SSE responses work in SillyTavern.
Can I use long character cards?
Yes, up to the model's context limit. 512K–2M models handle detailed cards, at a higher per-turn input cost.
Which model is best for roleplay?
Uncensored or Heretic variants. DeepSeek V4.1 Flash Uncensored is the strongest; Chimerax Heretic is the cheapest.
Can I run several characters at once?
Yes, but each session bills separately. Use a dedicated key per session if you want clean spend reports.
Why does quality drift after 50 turns?
The model is reading a long history and the oldest turns dominate. Summarize or branch a new chat; switching models rarely fixes it.
Can I use it on a phone?
Yes. SillyTavern is a web UI; the endpoint does not care about the device. Keep the key out of screenshots.
Next
Open the SillyTavern setup guide, then pick a model from the roleplay list. If you also write fiction outside chat, see the uncensored API guide for the one-shot workflow.
Key and session hygiene
- One key per device or per long-running session. Spend reports stay readable and revoking a leaked key does not kill everything.
- Set expiry. A roleplay key does not need to live for a year.
- Cap spend per key. A forgotten tab that keeps generating is the usual bill spike.
- Rotate after sharing logs or screenshots. Keys appear in more places than people expect.
A short prompt template that survives
Long system prompts are permanent input cost. Keep the instruction to what the card cannot carry itself:
Stay in character. Keep replies to 2–4 paragraphs.
Do not narrate the user's actions. End on an open beat.Put lore in the card or world info, not in the system prompt. The system block is resent every turn, so every extra sentence is billed on every turn of every session.
When replies get repetitive
Repetition usually means temperature is too low or the history is dominated by one phrasing. Raise temperature a little, then branch a new chat. Switching to a bigger model rarely fixes repetition and always costs more.