Janitor AI API setup: the exact proxy config, field by field

Quick answer: the five values JanitorAI asks for

JanitorAI's proxy setup has exactly five fields, and four of them are copy-paste. Open any chat, go to API Settings, pick Proxy as the API type, then create a new configuration with these values:

Configuration name:  APIsRouter-deepseek        (any label, only you see it)
Model name:          deepseek-v4-flash       (exact model ID, no spaces)
Proxy URL:           https://api.apisrouter.com/v1/chat/completions
API key:             sk-...                  (paste with no leading/trailing spaces)
Custom prompt:       (optional, leave blank to use the character card as written)

What each field actually does

Keep the custom prompt short and mechanical. Good uses: capping response length, enforcing third person, banning actions written for your character. Bad uses: pasting a second character sheet or a thousand-word style guide, which inflates every request and pushes older chat history out of context sooner. After saving, use the check proxy button if your build shows one, then send a short test message. If nothing comes back, work through the error table below before changing models.

• Configuration name: a local label. It is never sent to the API. Name it after the model, like APIsRouter-deepseek or APIsRouter-glm, so you can keep one saved config per model and switch in two taps.
• Model name: sent as the model parameter on every request. It must be the exact ID, lowercase with dashes: deepseek-v4-flash works, "DeepSeek V4 Flash" does not.
• Proxy URL: the full chat completions endpoint. JanitorAI sends requests to exactly the URL you type. SillyTavern takes a base URL ending at /v1 and appends the path itself, and it auto-populates the model list from /v1/models. JanitorAI does neither, which is why a URL that stops at /v1 returns a 404 here even though it works fine in SillyTavern.
• API key: sent as an Authorization: Bearer header to the proxy URL only. Never paste it into a chat message or screenshot.
• Custom prompt: an optional system-level instruction added alongside the character card, not a replacement for it. It is resent with every message, so every word costs input tokens for the whole life of the chat.

Which model name to enter, and why

The proxy URL and API key stay identical across configs. The only value that changes per model is the model name, so the practical move is one saved configuration per model you actually use. For roleplay, the DeepSeek V4 family is the approximate community favorite on JanitorAI boards: strong character-card adherence at the lowest token prices in this table. GLM-5, Kimi K2.6, and MiMo are solid value picks with policies that accommodate fictional content. Grok 4.5 is a common pick for darker or mature fictional themes because xAI's published policy explicitly allows them in fiction. Claude Sonnet 4.6 and Opus 4.7 write the best prose of anything here, but Anthropic's usage policy prohibits explicit content, so treat Claude as a strictly SFW creative-writing option. Two ground rules regardless of model. JanitorAI is an 18+ platform and its ToS still applies in proxy mode, and so does the content policy of whichever model provider you route to. Picking a model with a flexible fiction policy does not suspend either one.

Test the key outside JanitorAI first

When a config fails, you have three suspects: the URL, the key, and the model name. JanitorAI's error surface is thin, so test all three at once with curl before touching the app again:

curl https://api.apisrouter.com/v1/chat/completions \
  -H "Authorization: Bearer sk-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-flash",
    "messages": [{"role": "user", "content": "Say ready."}],
    "max_tokens": 20
  }'

Fixing 401, 404, and model not found

If the curl test returns a JSON completion, your URL, key, and model name are all valid, and the remaining problem lives inside JanitorAI: usually an unsaved config, a stale selected configuration, or a browser extension blocking the request. If curl fails, the status code tells you which field to fix.

What a month of chatting actually costs

JanitorAI resends the character card plus recent chat history with every message, so input tokens dominate the bill. A typical setup with a detailed card runs roughly 4,000 to 6,000 input tokens and 200 to 400 output tokens per message. The table assumes 5,000 in and 300 out across 1,000 messages, which is a heavy month for most users. Treat these as estimates: your card size and history depth move the numbers. These are APIsRouter catalog rates. APIsRouter is an OpenAI-compatible gateway with pay-as-you-go billing and no subscription; global models are listed at 20% off official pricing and Chinese models below official rates. There is no signup form: pay at the top-up page, the key arrives by email, and your first top-up adds a 100% balance bonus. That makes a $5 test of DeepSeek or GLM a months-long experiment rather than a monthly commitment.

FAQ