SillyTavern LLM settings for an OpenAI-compatible API

• SillyTavern installed and running. The steps follow the release branch.
• A RelaySwitch key from the console (email code sign-in). Keys start with sk-relay-.
• A balance. Top-ups start at $10, by card or crypto, with no subscription. Before the first top-up the console chat runs 12 free turns a day, so you can try a model before connecting it here.

1. Open API Connections (the plug icon in the top bar).
2. Set API to Chat Completion.
3. Set Chat Completion Source to Custom (OpenAI-compatible).
4. Custom Endpoint (Base URL): https://api.relayswitch.io/v1. Keep the /v1. SillyTavern adds the rest; its docs put it plainly: "Do NOT add the /chat/completions suffix."
5. Custom API Key: your key. SillyTavern hides it after you connect.
6. Click Connect. The Available Models dropdown fills from /v1/models. Pick a model there, or type its id in Enter a Model ID.
7. Click Test Message. It sends one short request, which bills like any other.

Use the id the Models page prints, such as claude-sonnet-5-5, not the display name.

SillyTavern can also send the Anthropic Messages format, through its Claude source. The reverse proxy block sits in the same connection panel; SillyTavern's docs do not cover it, so the labels here are from the UI.

1. Set Chat Completion Source to Claude.
2. Proxy Server URL: https://api.relayswitch.io/v1. SillyTavern adds /messages.
3. Proxy Password: your sk-relay- key. The field's note says it is used as the password for the proxy instead of an API key, and the Claude API key field above it is ignored once a proxy is set.
4. Pick the model in Claude Model, then click Connect.

Claude Model is a fixed dropdown. These 9 of its ids answer the Anthropic format in the catalog today:

The dropdown has no entry for claude-fable-5.1, claude-haiku-5-5, claude-opus-5-5, claude-sonnet-5-5. Reach those through the Custom source, where any id can be typed.

Prompt Post-Processing, in the connection panel under the Custom source, reshapes the message list before it leaves SillyTavern. The options, as the panel lists them:

• None
• Merge consecutive roles (with tools)
• Semi-strict (alternating roles; with tools)
• Strict (user first, alternating roles; with tools)
• Merge consecutive roles (no tools)
• Semi-strict (alternating roles; no tools)
• Strict (user first, alternating roles; no tools)
• Single user message (no tools)

Leave it on None unless a model complains. The merge options join back-to-back messages from the same role; Semi-strict also keeps user and assistant turns alternating. Pick Strict when a model answers a 400 about roles or message order: a character chat opens on the card's greeting, an assistant message, and Strict puts a user message first. The "no tools" variants drop function calling; keep a "with tools" one if you use tool calls.

With a Custom source, the AI Response Configuration panel shows Temperature, Frequency Penalty, Presence Penalty, Top P, Seed and Max Response Length (tokens). Those are the standard Chat Completions parameters. SillyTavern's server builds this body for every message:

Custom stopping strings add a stop list, a Seed of 0 or more adds seed, and Multiple swipes per generation adds n. The API forwards all of these to the model unchanged. A model that refuses one answers with a 400 error.

Top K, Min P, Top A, Repetition Penalty, DRY, XTC, Mirostat and the other Text Completion samplers do not appear for a Custom source, and SillyTavern does not send them. Settings you made for a local backend have no effect here. To send one anyway, open Additional Parameters in the connection panel and add it under Include Body Parameters as YAML:

The API passes extra fields through as well. Whether a model applies min_p or repetition_penalty depends on the model; the four standard samplers are the ones to tune first. To stop SillyTavern sending a field a model rejects, list it under Exclude Body Parameters:

Starting points for long-form fiction: Temperature 0.8 to 1.0, Top P 0.95, Frequency Penalty 0.1 to 0.3 against repeated phrases, Presence Penalty 0 to 0.3 to push new topics. Change one at a time and swipe a few replies before judging.

Context Size (tokens) is the most SillyTavern sends as the prompt: card, system prompt and chat history, minus the response length. A fresh install starts Chat Completion at 4,095 tokens of context and 300 response tokens. On a hosted model with a far larger window, that drops older messages long before the model needs it to.

SillyTavern does not read a context window from a Custom source. The slider goes past any model's limit, so set it from the Context column in the table below. Set it too high and a long chat fails once the prompt passes the model's window; Euryale 70B v2.3 stops at 20K tokens.

Context Size is also the cost dial. Once a chat fills it, every message sends that many input tokens, so a 32K setting costs several times more per message than a short chat on the same model. The costed example below prices both.

Request model reasoning in the AI Response Configuration panel shows a model's reasoning as a collapsible block above the reply when the model returns it in a separate field. For a model that writes its reasoning inline between <think> tags, turn on Auto-Parse under Advanced Formatting > Reasoning instead.

Reasoning Effort on a Custom source is sent only for model ids SillyTavern knows from OpenAI, such as gpt-5.5. For any other model, put reasoning_effort: high (or low, medium) under Include Body Parameters. The API passes it to the model, and models with an effort setting use it.

Reasoning tokens are output tokens. They count against Max Response Length and bill at the output rate, even when the block stays collapsed. If replies come back empty or cut off, raise Max Response Length; SillyTavern's docs suggest 1024 to 4096 for reasoning models.

Four general models and 2 the catalog tags for fiction and character dialogue (Gemma 4 31B Fabled, Euryale 70B v2.3). Prices are per million tokens. The message columns assume 3,000 tokens in (card plus recent history) and 300 out, the default response length.

The general models carry the largest context windows. The fiction-tuned builds are trained on stories and dialogue and have smaller windows, so check the Context column before you raise Context Size. The roleplay chat runs character cards in the browser on the same key, and the character card guide explains what each card field costs in tokens.

Early in a chat on Claude Sonnet 5.5, a message sends about 3,000 tokens and gets 300 back: $0.0045 each, or $0.45 for 100 messages.

Later, with Context Size at 32K and the history filling it, each message sends 32,000 tokens: $0.0393 each, or $3.93 for 100. The same 100 full-context messages on GLM 5.3 come to $1.08.

Swipes, Continue and Test Message are requests too, each billed the same way. When the model reports cached input tokens, they bill at its cached rate, which lowers the cost of a long chat that resends the same history. The console's Usage page lists every request with its tokens and charge. The pricing page has the full billing rules.

List the models your key can call, from the machine that runs SillyTavern:

• 401: the key is wrong or revoked. Paste it again in Custom API Key and click Connect.
• 402: the balance is too low. Top up in the console; the next message goes through.
• 404: the model id is not in the catalog. Pick it from Available Models instead of typing it.
• 429: the key hit its rate limit. Wait a few seconds before the next swipe.
• A 400 naming a parameter: the model refuses that field. Add it to Exclude Body Parameters, or remove it from Include Body Parameters.
• A 400 about roles or message order: some models want alternating user and assistant turns. Set Prompt Post-Processing in the connection panel to Semi-strict, or Strict if the first message must be the user's.
• Connection fails or the model list stays empty: the URL is missing /v1 or ends in /chat/completions. Use https://api.relayswitch.io/v1 exactly.
• Empty or cut-off replies: a reasoning model used up Max Response Length before it wrote the reply. Raise it.
• Old messages forgotten: Context Size is still at 4,095. Raise it toward the model's window.

The API docs list every error code.

Menu paths and field names are from SillyTavern's docs, read 2026-10-27: Chat Completions, Custom OpenAI-compatible endpoint, Common Settings and Reasoning. The request body, the default context and response sizes, and which samplers show for a Custom source are from SillyTavern's source on the release branch, read the same day. The Claude source's reverse proxy block, its Claude Model dropdown and the Prompt Post-Processing options are from SillyTavern's UI on the release branch, read 2026-11-02. For a chat client with a model allowlist and per-user keys, see the Open WebUI guide.