Connect SillyTavern to NezhaGate: 10-Minute Setup
A beginner-proof walkthrough: point SillyTavern at NezhaGate, pick a model and send your first message. Includes model picks and fixes for the usual errors.
About 10 minutes · Written against SillyTavern 1.19
What you need
- A computer running Windows, macOS or Linux.
- A NezhaGate account and API key (step 1 shows you how).
- About 10 minutes. Already have SillyTavern installed? Skip to step 3.
Step 1: Sign up for NezhaGate and create an API key
- Open nezhagate.com and click Sign up free in the top-right corner. Email works, and so does one-click sign-in with Google or GitHub.
- Once you are in, go to Console → API Keys. Under Create API Key, give the key a name (for example
sillytavern), leave Per-Key Limit at 0 for no limit, and click Create Key. - A key starting with
sk-appears. Copy it right away: it is shown only once. Paste it into a notepad for the next steps.
[Screenshot: Console → API Keys, highlighting the Create Key button and the copy button of the new key]
Check and top up your balance on the Billing page. You pay only for what you use, and failed requests are never billed. Per-model rates are on the pricing page.
Step 2: Install and start SillyTavern
SillyTavern needs two free tools first: Node.js (version 20 or newer; download the LTS installer and click through it) and Git.
Windows
- Create an ordinary folder such as
D:\AI. Do not use a system folder like Program Files or System32. - Open that folder, type
cmdinto File Explorer's address bar and press Enter. A black command window opens. - Paste this line, press Enter and wait for the download to finish:
git clone https://github.com/SillyTavern/SillyTavern -b release
- Open the new
SillyTavernfolder and double-clickStart.bat(do not choose "Run as administrator"). The first start installs its dependencies, so give it a minute.
macOS / Linux
Open Terminal, paste this line and press Enter:
git clone https://github.com/SillyTavern/SillyTavern -b release && cd SillyTavern && ./start.sh
When it is up, your browser opens http://127.0.0.1:8000 by itself. Next time, just run Start.bat (or ./start.sh) again, and keep that command window open while you use SillyTavern.
[Screenshot: SillyTavern's main screen on first launch, highlighting the plug icon in the top bar]
Step 3: Connect SillyTavern to NezhaGate
- Click the plug icon in the top bar (its tooltip says API Connections).
- In the API dropdown, choose Chat Completion.
- Set Chat Completion Source to Custom (OpenAI-compatible).
- In Custom Endpoint (Base URL), enter:
https://nezhagate.com/v1
- Paste the key from step 1 into Custom API Key.
- Click Connect. When the status reads Valid, the address and key are right, and the Available Models list fills in.
- Pick a model under Available Models;
deepseek-v4.1-flashis a good first choice. You can also type a model ID into Enter a Model ID. - Click Test Message. API connection successful! means you are done. The test is one real call and costs a fraction of a cent.
[Screenshot: the API Connections panel with Chat Completion Source, Custom Endpoint, API key, Connect, Available Models and the Valid status marked in order]
Stop the address at
/v1. SillyTavern adds /chat/completions on its own; adding it yourself gives a 404.Now go back to the chat screen, pick a character card and start talking.
Step 4: Change two settings (important)
Click the sliders icon at the far left of the top bar to open AI Response Configuration, and change these two values. SillyTavern's defaults are sized for small models and cause trouble with modern ones:
| Setting | Default | Use | Why |
|---|---|---|---|
| Max Response Length (tokens) | 300 | 2000–4000 | DeepSeek, Kimi, GLM, Qwen and Grok think before they answer, and the thinking counts toward this length. At 300, replies get cut off or come back empty. |
| Context Size (tokens) | 4095 | 16000–32000 | How much chat history goes with each message. Too small and characters forget; too large and every message costs more. |
You are billed only for what the model actually generates, so a higher Max Response Length does not cost more by itself. It is just a ceiling.
Recommended models
| You want | Model ID | Notes |
|---|---|---|
| Cheap and fast, for long campaigns | deepseek-v4.1-flash | Low price, natural English and Chinese. Turn thinking off for faster, cheaper replies (see below). |
| The best prose, narrative, long form | claude-fable-5 | Keeps character motivation and emotional arcs coherent; made for fiction and script-style role-play. Priced higher. |
| Big lorebooks and heavy character cards | kimi-k3 | Strong long-context understanding. Its temperature is fixed by the model, so the Temperature slider has no effect. |
| A good all-rounder | glm-5.3, qwen3.7-max | Both always think first: keep Max Response Length at 2000 or more. |
Current prices are on the pricing page. To switch models, pick another one under Available Models.
Turn off DeepSeek's thinking
In the API Connections panel, click Additional Parameters, put these two lines into Include Body Parameters and save:
thinking: type: disabled
Remove them again when you switch to another model.
Common problems and errors
| What you see | What to do |
|---|---|
| "Status check bypassed" after Connect, and the model list is empty | The address or key is wrong, or the network cannot reach us. Check that the address is https://nezhagate.com/v1 and that the key was copied whole, with no spaces around it. |
| Test Message says "Could not get a reply from API" | Same checks as above, then make sure your account has a balance. |
| 404 | The address ends in /chat/completions. Delete that part and keep /v1. |
| 401 | The key is wrong or was deleted. Create a new one in the console and copy it. |
| 402 | Your balance ran out, or the key hit the limit you set. Top up on the Billing page or raise the key's limit. |
400 about an unsupported parameter (for example stop with grok-4.7) | Click Additional Parameters and add a line - stop to Exclude Body Parameters (use the parameter named in the error). |
| Replies stop halfway or come back empty | Raise Max Response Length (step 4). |
| Changing Temperature seems to do nothing | Some models (Claude 5 generation, Kimi K3, the GPT family) ignore sampling settings. That is expected and causes no error. |
No API key yet?
Sign up and create a key in the console. Pay as you go; failed requests are never billed.
Get API Key → PricingMore guides: Cherry Studio · Cline
