A Janitor AI proxy is not a VPN: it's an API connection that sends your chat to a bigger model than Janitor's own JLLM. You set it up in a chat's API Settings with three things, a model name, a proxy URL and an API key, and OpenRouter is the route Janitor's own help centre walks you through. On OpenRouter's list on 14 November 2026, 100 messages cost from about 3 cents (Qwen3.7 Flash, GLM 5.3 Flash) to about $2 (Claude Sonnet 5.5) at an 8,000-token chat, and the free model Janitor's quickstart still names no longer exists.
If what you really want is a story that stays straight past chapter three, there's a shorter route at the end: a writing app with the AI already built in.
What a Janitor AI proxy is, and why people use one
Janitor's own model, JLLM, is free but small in memory: about 8k–9k tokens, as we found in our Janitor AI alternatives check. A proxy swaps it for any model that speaks the OpenAI chat format. Seven of the 13 models priced below take a million tokens, so the bot can hold a very long chat in view, and you pay that model's provider per message.
What people search around it, from Google Keyword Planner (US, monthly, pulled 14 November 2026):
| Search | Monthly searches | Trend (3 months vs 3 before) |
|---|---|---|
| janitor ai proxy | 2,400 | ×0.82 |
| janitor ai alternative | 2,400 | ×1.22 |
| best janitor ai proxy | 260 | ×1.07 |
| free janitor ai proxy | 140 | ×0.62 |
| how to use openrouter on janitor ai | 90 | ×0.90 |
| how to use janitor ai proxy | 90 | ×0.62 |
| openrouter for janitor ai | 50 | ×1.20 |
How to set up a proxy on Janitor AI
Janitor's TLDR Quickstart Proxy Instructions (updated 01/08/2025) give the steps:
- Make an account at OpenRouter and create a key under Keys. Treat it like a password.
- Open a chat on Janitor, click the gear at the top right, choose API Settings, then Add Configuration.
- Fill in Model Name, Proxy URL (
https://openrouter.ai/api/v1/chat/completions) and API Key. Leave Custom Prompt empty, as Janitor's guide says. This page doesn't hand out jailbreak prompts. - Click Add, then Save Settings, refresh the page and send a test message.
Paste the full URL, /chat/completions included. Janitor's OpenRouter Error Guide (updated 04/08/2025) lists "A network error occurred" as the result of a wrong URL or of not refreshing after you save.
Checked 14 November 2026: the quickstart's model name, deepseek/deepseek-chat-v3-0324:free, is not in OpenRouter's model list. The paid deepseek/deepseek-chat-v3-0324 is still there at $0.29 per million input tokens and $1.14 per million output. Of 471 models listed that day, 14 had a :free version, and none of them was a DeepSeek model. Copy the model ID from OpenRouter's site on the day you set it up, never from a guide.
What 100 messages cost on each proxy model
Every message sends the bot's definition and the chat so far, so the input is what you pay for. We priced two chats from OpenRouter's listed prices on 14 November 2026: 8,000 tokens in (about JLLM's window) and 32,000 tokens in (a long chat on a big model), each with a 350-token reply (about 260 words). Prompt caching, which some hosts discount, isn't counted, so a real bill can be lower.
| Model (OpenRouter ID) | $ per 1M in / out | Context | 100 messages at 8k | 100 messages at 32k | Our studies |
|---|---|---|---|---|---|
Qwen3.7 Flash (qwen/qwen3.7-flash) | 0.03 / 0.13 | 1M | $0.03 | $0.10 | — |
GLM 5.3 Flash (z-ai/glm-5.3-flash) | 0.04 / 0.08 | 1M | $0.04 | $0.13 | — |
DeepSeek V4.1 Flash (deepseek/deepseek-v4.1-flash) | 0.05 / 0.20 | 1M | $0.05 | $0.17 | Wrote 21 of 21 scenes on sexual themes (non-explicit); 0.7 stock phrases per 1,000 words |
DeepSeek V4 Flash (deepseek/deepseek-v4-flash) | 0.09 / 0.18 | 1M | $0.08 | $0.29 | — |
Mistral Small 4 (mistralai/mistral-small-2603) | 0.15 / 0.60 | 262k | $0.14 | $0.50 | — |
DeepSeek V3.2 (deepseek/deepseek-v3.2) | 0.26 / 0.38 | 164k | $0.22 | $0.85 | — |
DeepSeek V3 0324 (deepseek/deepseek-chat-v3-0324) | 0.29 / 1.14 | 164k | $0.27 | $0.97 | — |
Llama 4 Maverick (meta-llama/llama-4-maverick) | 0.35 / 1.00 | 524k | $0.32 | $1.16 | 5.0 stock phrases per 1,000 words |
Gemini 3.5 Flash Lite (google/gemini-3.5-flash-lite) | 0.30 / 2.50 | 1M | $0.33 | $1.05 | — |
Kimi K2.6 (moonshotai/kimi-k2.6) | 0.65 / 3.41 | 262k | $0.64 | $2.20 | — |
Grok 4.3 (x-ai/grok-4.3) | 1.25 / 2.50 | 1M | $1.09 | $4.09 | — |
Grok 4.7 (x-ai/grok-4.7) | 2.00 / 6.00 | 500k | $1.81 | $6.61 | Wrote 34 of 34 explicit love scenes asked; 0 stock phrases per 1,000 words |
Claude Sonnet 5.5 (anthropic/claude-sonnet-5.5) | 2.00 / 10.00 | 1M | $1.95 | $6.75 | 31 of 36 explicit, 5 softened; 0 stock phrases per 1,000 words |
Source: prices and context from OpenRouter's models API, 14 November 2026; cost per 100 messages is our arithmetic. "Our studies" is KudoWrite's Fade to Black (consenting adult characters, rated by heat) and AI Slop Index (stock phrases such as "voice barely above a whisper"), both 22 October 2026, and Banned by the Bot (21 October 2026); a dash means we didn't test that model. Full rows: the data file.
At 8k the spread is about 65 to 1: $6.75 buys 100 long messages on Claude Sonnet 5.5, or about 4,000 on DeepSeek V4.1 Flash. For romance at a price you can live with, DeepSeek V4.1 Flash is the cheap model we have data on: in our Banned by the Bot study it wrote every scene on a sexual theme we asked for, though those were non-explicit, so its open-door heat is untested. Grok 4.7 is the one we measured writing explicit love scenes every time, at nearly 40 times the price.
DeepSeek's own API is the other route people use. On DeepSeek's pricing page (14 November 2026), deepseek-flash costs $0.15 in and $0.60 out per million tokens off-peak, double that at peak hours (01:00–04:00 and 06:00–10:00 UTC on weekdays): $0.14 per 100 messages at 8k off-peak, $0.28 at peak. That's more than OpenRouter's listed price for the same family that day.
Is there a free Janitor AI proxy?
Yes, with a ration. OpenRouter's limits page allows 20 requests a minute on :free models and 50 a day, or 1,000 a day once you've bought at least 10 credits. Janitor's quickstart words it the same way: 50 messages a day, resetting at midnight UTC, or "add $10 once for 1000/day." Its error guide adds that a failed message counts toward the 50.
Free variants change without notice. The DeepSeek one Janitor's guide was built around is gone, and the error guide's fix for "No endpoints found for model name" is to move to another version. Its fix for "No endpoints found matching your data policy" is to switch on Model Training in your OpenRouter privacy settings, which lets your chats be used to train models.
Community and vendor proxies (MegaNova, NanoGPT, Chutes, the "Gemini for Janitors" relays) fill the gap. MegaNova's own setup guide (16 June 2026) gives no usage caps and quotes Janitor's notice that it isn't affiliated with proxy services and that the proxy provider may log your IP and chats.
Why your Janitor proxy isn't working
From Janitor's OpenRouter Error Guide, the ones you'll meet most:
| Error | What it means |
|---|---|
| 429 "free-models-per-day" | You've used today's 50 free messages; errors count too |
| 429 "temporarily rate-limited upstream" | That model is overloaded; pick another or its paid version |
| 402 "Prompt tokens limit" | Your OpenRouter balance is too low for that chat's size |
| 404 "No endpoints found for model name" | The model ID was retired or is down |
| 400 "maximum context length" | Bot definition, memory and chat are bigger than the model takes |
| "A network error occurred" | Wrong proxy URL, or you didn't refresh after saving |
NanoGPT's Janitor docs add one of their own: use nano-gpt.com, not api.nano-gpt.com, because Janitor's servers can't reach the api. address.
Where the other options fall short
- JLLM: free, but about 8k–9k tokens of memory, so a long chat loses its start.
- OpenRouter's free models: 50 messages a day until you pay $10, the model IDs come and go, and the error guide's way past one data-policy error is letting your chats train models.
- Chutes: RoboRhythms (20 June 2026, updated 19 September) reports it raised DeepSeek 3.2 from $0.24/$0.42 to $1.00/$1.00 per million tokens with no announcement, retired R1-0528 and dropped its one-time payment.
- Subscriptions: the same article puts NanoGPT at $8 a month and Featherless at $10–25 a month, with the $10 tier capped at a 16,000-token context, billed whether you chat that month or not.
- Every proxy: nothing files who the characters are or what happened between them; you rely on what still fits in the window.
A story instead of a proxy: KudoWrite
| What you set up | What you pay | What keeps track of the characters | |
|---|---|---|---|
| KudoWrite | Nothing: the AI is built in | One free credit, then one-off packs from $3.50 for 100 credits that never expire | A story bible that updates itself as you write |
| Janitor + OpenRouter free | Key, URL, model ID | 50 messages a day, or $10 for 1,000 | The model's window |
| Janitor + paid proxy | Key, URL, model ID | Per message, $0.03–$6.75 per 100 above | The model's window |
| Janitor + NanoGPT / Featherless | Account, URL, model ID | $8–25 a month | The model's window |
KudoWrite is a writing app where the chat you'd want to keep becomes a book. You type a line, the AI Expands it into prose in place, and the story bible files your characters, places and what happened between them as you go, so the vampire still has the scar from chapter one when you reach chapter twelve. There's no proxy URL, no key and no model ID to break: KudoAI runs on a current DeepSeek "flash" model, routed only to zero-data-retention hosts that don't train on prompts, and KudoWrite adds no content filter of its own. Paste your best chat log into a .txt or .docx, import it, and its chapters are detected for free. Then start the book and keep going from where the bot lost the thread.
For which model writes best once you're past Janitor, see the best OpenRouter models for creative writing and what's actually free on OpenRouter.