Connect Janitor AI or SillyTavern
This page lists the exact settings we tested for using an API key from this service in two chat front ends. Last checked 2026-11-02.
Use is subject to the Usage Rules. Home
What you need
- An API key from your account. The examples below use the placeholder
YOUR_API_KEY: replace it with your own key and never paste the real key into chats, screenshots or posts. - The values from the table.
| Field | Value |
|---|---|
| Proxy URL (Janitor AI) | https://api.privatedeskai.com/v1/chat/completions |
| Custom endpoint (SillyTavern) | https://api.privatedeskai.com/v1 |
| API key | YOUR_API_KEY |
| Model, first choice | deepseek-v4-flash-fast, response limit (max tokens) 400 |
| Model, second choice (reasoning on) | deepseek-ai/DeepSeek-V4-Flash-0731, response limit (max tokens) 1000 |
Janitor AI
- Open a chat, open the menu at the top of the chat and choose API Settings.
- Open the Proxy tab and click +Add Configuration.
- Fill in the fields: Config Name (any name you like); Model Name
deepseek-ai/DeepSeek-V4-Flash-0731, typed exactly; Proxy URLhttps://api.privatedeskai.com/v1/chat/completions; API KeyYOUR_API_KEY(replace it); leave Custom Prompt empty. - Click Save Settings and reload the page. Do not rely on the Check Key/Model button: Janitor's own help suggests ignoring it.
- Send a message. If replies stop in the middle of a sentence, raise the response length (max tokens) in Janitor's generation settings to 1000 for this model.
What we verified. On 2 November 2026 the project owner connected Janitor AI in a desktop browser with exactly the model name and Proxy URL above. The request went through and a character reply arrived in about 14.07 seconds with a character description of about 37,000 tokens. This was one reply, not a load test, and it used the model name shown above: we have not yet tried the deepseek-v4-flash-fast name inside Janitor itself.
SillyTavern
- Open API Connections and choose Chat Completion as the API.
- Set Chat Completion Source to Custom (OpenAI-compatible).
- Enter
https://api.privatedeskai.com/v1as the custom endpoint andYOUR_API_KEY(replace it) as the custom API key. - Enter the model name from the table,
deepseek-v4-flash-fastordeepseek-ai/DeepSeek-V4-Flash-0731, and connect. - In the response settings set the maximum response length (tokens) to the value that matches the model: 400 for
deepseek-v4-flash-fast, 1000 fordeepseek-ai/DeepSeek-V4-Flash-0731. Streaming can stay on, which is the SillyTavern default.
What we verified. We ran a 10-message role-play scene against the server of SillyTavern 1.19.0 (the official release image) with this source type and these values. The test called the server API that the interface itself uses; it did not click through the web interface, so field labels can differ a little between versions. Results are on the reliability page.
Why two model names and two limits
deepseek-v4-flash-fast is an alias of the same model with reasoning switched off, at the same price per token. The regular model reasons before it answers: in our 220-request test it returned hidden reasoning in 138 of 220 replies. Those reasoning tokens use up the response limit and are billed like any other output. With the regular model at a limit of 400, 18 of 220 replies (8.2 percent) ended at the limit and were cut off mid-sentence, which is why we suggest a limit of 1000 for that model.
Next
See what a message costs and how that compares with other options on the prices and comparison page, and read what we measured and where it falls short on the reliability page. Usage rules: acceptable use and terms of service.