The best Janitor AI API settings for 2026: Temperature 0.9, Max tokens 0 for reasoning models (800–1000 for others), Context size 32k. Below: what every setting does, a copy-paste roleplay prompt, and fixes for short, repetitive or cut-off replies.

WHERE ARE THE API SETTINGS?
Open any chat → click "using janitor" (or the top-right menu) → API Settings. Two Provider choices:
• JanitorLLM: free and built in, but small memory and repetitive replies.
• Proxy: connect a real model (DeepSeek, Claude, Gemini, GPT, GLM...) through a provider like NanoGPT. Much better roleplay, longer memory.

QUICK PROXY SETUP (NANOGPT)
1. Get a key at nano-gpt.com/api and add a small balance. No account needed.
2. API Settings → Proxy → + Add.
3. Proxy URL: https://api.nano-gpt.com/api/v1/chat/completions
4. Paste your API key. Model: auto-model-basic
5. Apply, set it Active, refresh the page.
Full walkthrough: Best Proxy for Janitor AI (2026)

MAIN SETTINGS

Temperature → 0.8 to 1.0
How creative the model is. Too high and it rambles or breaks character. Start at 0.9; lower it if replies get weird, raise it if they feel flat.

Max tokens → 0 for reasoning models, 800–1000 for others
The longest a reply can be. Reasoning models' thinking counts toward this limit, so set 0 (no limit) or replies get cut off.

Context size → 32k
How much of the chat the model remembers. Bigger = better memory but higher cost, since every message re-sends it. Go 64k+ for long lore-heavy stories, never above your model's limit.

ADVANCED SETTINGS
Not sure? Leave at 0 and the model uses its defaults.
• Top P → 0.9–0.95: trims unlikely word choices.
• Top K → 0 (off) or 40: most people leave it off.
• Repetition penalty → 1.05–1.1: stops repeated words in a reply. 0 for reasoning models.
• Frequency penalty → 0.1–0.3: stops recycled phrases across the chat. 0 for reasoning models.

COPY-PASTE ROLEPLAY PROMPT
Paste into the Global Prompt box:

You are {{char}} in an ongoing roleplay with {{user}}. Stay in character at all times and never speak or act for {{user}}. Write vivid, immersive replies of 2–4 paragraphs that move the story forward, react to what {{user}} just did, and keep {{char}}'s personality, voice and memories consistent. Show emotion through actions, body language and dialogue. Vary your sentence openings and avoid repeating earlier phrases. End each reply where {{user}} can naturally respond.

PRESETS BY MODEL TYPE
Reasoning models: Temperature 0.8–1.0 • Max tokens 0 • Context 32k • Penalties 0
Regular models: Temperature 0.9 • Max tokens 800–1000 • Context 32k • Top P 0.95 • Repetition 1.05 • Frequency 0.2
Want reasoning visible in chat? Proxy URL: https://api.nano-gpt.com/api/v1thinking/chat/completions

TROUBLESHOOTING
Too short: raise Max tokens; add "Write 3–4 paragraphs" to your prompt.
Cut off mid-sentence: Max tokens too low; use 0 for reasoning models.
Repeats itself: Repetition penalty 1.1, Frequency penalty 0.3, or start a fresh chat.
Incoherent: lower Temperature to 0.7–0.8.
Forgets things: raise Context size.
Reasoning shows in chat: switch from the v1thinking URL to the regular /v1 URL.
Speaks for you: use the prompt above.
Costs adding up: lower Context to 16–32k or use a cheaper model.

MORE HELP
Quick 5-step setup: Janitor AI API Tutorial
Full proxy guide: Best Proxy for Janitor AI (2026)
Models and prices: nano-gpt.com/pricing
Community: NanoGPT Discord