Checked 2026-11-16

Janitor AI proxy: what it is, the settings, the models, the price

A proxy on Janitor AI is the Provider setting that sends each message to a chat completions URL you choose, on a key you pay for per token. This page has the exact fields, the builds that answer a character card as written, Claude and DeepSeek by proxy, what happens when a model goes down, and 1,000 messages priced from the live catalog.

Minimum top-up
$10
Pay with
USDT, USDC, BTC, ETH, SOL, card or WeChat Pay
Sign-up
An email address, or sign in with Google or GitHub. No phone number, no identity documents.
Models
289 on one key, prices read live
Base URL
https://api.relayswitch.io/v1OpenAI Chat Completions and Responses on 280 chat models, Anthropic Messages on 280, Gemini generateContent on 280
One message, example
$0.00031 on Manta Flash 1.0; $0.0033 on Claude Sonnet 5.5 for one 2K-in / 300-out turn
Subscription
None. Prepaid credit, billed per token.

In Janitor AI, a proxy is not a VPN and not a server you run. It is the Proxy option under Provider in a chat's API Settings: a Proxy URL, an API key and a Model Name. With it saved, Janitor stops using its own model and posts every message of the chat, card and persona included, to that URL, and the reply comes from the model you named. You pay the API behind the URL per token; Janitor itself charges nothing for the proxy.

RelaySwitch is one such API. The Proxy URL is https://api.relayswitch.io/v1/chat/completions, the key starts with sk-relay-, and Model Name takes any chat id on the Models page: the uncensored and abliterated builds that answer a card without a jailbreak, every Claude id, DeepSeek, Grok, GLM and GPT. Keys are prepaid from $10 by card or crypto, no KYC and no subscription needed, and every reply carries a signed receipt. Prices on this page are read from the catalog when it is built, never typed.

01What proxy means in Janitor AI

Janitor AI ships with its own model and sells access to it. The proxy setting exists so that a chat can use any other model instead, through an OpenAI-compatible chat completions endpoint. Three things follow from how Janitor built it:

  • Janitor relays the request from its own servers. The API behind the Proxy URL sees Janitor's address, not yours; your browser only talks to Janitor. A "network error" is therefore between Janitor and the API, not on your connection.
  • The character's creator decides whether proxies are allowed. A card marked proxy-off ignores your configuration; pick another card or allow proxies on one you made.
  • The proxy is per configuration, not per account. You can keep several (one per model, one per key) and switch the active one from the same panel. Generation settings such as max tokens and temperature live elsewhere and travel with every request unchanged.

A proxy is a "Janitor AI API key" in the sense people mean when they search for one: the key is the API's, not Janitor's. Janitor publishes no public API of its own; the proxy panel is where a key goes.

02The exact settings

  1. Create a key in the console (email code sign-in) and top it up. Narrow the key to a model list and give it a USD budget if you share the account or chat late at night.
  2. Open a chat whose creator allows proxies. Click the gear or menu icon at the top right and choose API Settings.
  3. Under Provider, pick Proxy, then + Add Configuration. Older layouts show the same panel behind the using janitor button above the chat.
  4. Fill the five fields exactly as below.
  5. Click Add, then Save Settings, then refresh the page. Open the configuration again and check it reads Active.
  6. Send a message. That is the test: Janitor's own quickstart calls the Check Key/Model button unreliable, and a Network Error after pressing it means nothing.
text
1Config Name: RelaySwitch
2Model Name: glm-5.3-flash-uncensored
3Proxy URL: https://api.relayswitch.io/v1/chat/completions
4API Key: sk-relay-...
5Custom Prompt: (leave blank)
  • Proxy URL is the full path. Janitor posts to the address as given, so a URL that stops at /v1 gets a 404. SillyTavern, Chub and Risu take the shorter https://api.relayswitch.io/v1 and add the rest themselves; Janitor does not.
  • Model Name is the id, pasted exactly. Janitor sends it as the model field, so a display name or a stray space is a 404 too.
  • Custom Prompt stays blank. The card and your persona already go with every message; a prompt here is billed as input on every turn, for the whole chat.

The JanitorAI setup guide walks the same fields with every error code Janitor prints; the proxy not working page is the table to open when a message fails.

03Which models work best for roleplay

Three kinds of model get typed into Janitor, and they differ more in behaviour than in price. The uncensored and abliterated builds answer the card as written: no refusal on a dark scene, no jailbreak prompt riding along as paid input. Claude writes the longest, most consistent prose and holds a persona across a long chat, at a higher price per message and with its own content rules. The other flagships sit in between. Every row below is live in the catalog today, with one 2K-in / 300-out turn and 1,000 of them priced at the current rate.

ModelModel NameKindContextInputOutputone 2K-in / 300-out turn1,000 messages
GLM-5.3 Flash Uncensoredglm-5.3-flash-uncensoreduncensored build—$0.1053$0.3685$0.00032$0.3211
GLM-5.3 Uncensoredglm-5.3-uncensoreduncensored build—$0.282$0.846$0.00082$0.8178
GLM-5.3 Abliteratedglm-5.3-abliterateduncensored build—$2.1825$4.365$0.0057$5.67
Kimi K3.1 Pro Uncensoredkimi-k3.1-pro-uncensoreduncensored build1M tokens$1.164$5.82$0.0041$4.07
DeepSeek V4.1 Flash 1017 Uncensoreddeepseek-v4.1-flash-1017-uncensoreduncensored build1M tokens$0.8003$3.201$0.0026$2.56
DeepSeek V4.1 Flash Abliterateddeepseek-v4.1-flash-abliterateduncensored build—$0.8488$1.6975$0.0022$2.21
DeepSeek V4 Pro Uncensoreddeepseek-v4-pro-uncensoreduncensored build—$0.97$1.94$0.0025$2.52
Apex Flash 1 Abliteratedapex-flash-1-abliterateduncensored build262K tokens$0.68$2.32$0.0021$2.06
Gemma 4 31B Fabledgemma-4-31b-fableduncensored build262K tokens$0.35$1.575$0.0012$1.17
Gemma 4 31B Novelistgemma-4-31b-novelistuncensored build262K tokens$0.35$1.575$0.0012$1.17
Cydonia 24B v4.3cydonia-24b-v4.3uncensored build131K tokens$0.484$0.60$0.0011$1.15
Uncensored AI Roleplayuncensored-ai-roleplayuncensored build128K tokens$0.60$2.34$0.0019$1.90
Claude Sonnet 5.5claude-sonnet-5-5Claude1M tokens$1.20$3.00$0.0033$3.30
Claude Opus 5.5claude-opus-5-5Claude1M tokens$2.716$9.70$0.0083$8.34
Claude Haiku 4.5claude-haiku-4-5Claude200K tokens$0.60$3.00$0.0021$2.10
Claude Fable 5claude-fable-5Claude1M tokens$3.00$15.00$0.0105$10.50
DeepSeek V4 Flashdeepseek-v4-flashflagship1M tokens$0.30$1.20$0.00096$0.96
Grok 4.6grok-4.6flagship500K tokens$0.60$1.80$0.0017$1.74
GLM 5.3glm-5.3flagship1M tokens$0.3178$0.9987$0.00094$0.9352
GPT-5.5gpt-5.5flagship1.1M tokens$1.80$9.00$0.0063$6.30

USD, read from the RelaySwitch catalog when this page was built. A message is 2,000 tokens in (card, persona and recent history) and 300 out; the last column is 1,000 such messages. Input and output are per million tokens.

The spread is wide: 1,000 messages run from $0.3211 on GLM-5.3 Flash Uncensored to $10.50 on Claude Fable 5. A build that thinks before it answers spends part of the reply budget on thinking; if replies come back short or empty, raise max tokens in Janitor's generation settings to 1,000 or more, or pick a build without a thinking step. The best models for JanitorAI page ranks the chat models by calls through this API in the last seven days, so you can see what other Janitor users actually type, and the uncensored models page lists every tagged build with its daily check. To try a model before pasting it into Janitor, the roleplay chat runs a character card on the same key in the browser.

04Claude proxy for Janitor AI

A Claude proxy for Janitor AI is the same configuration with a Claude id in Model Name. Janitor speaks the OpenAI chat completions format only, and every Claude id here answers that format at https://api.relayswitch.io/v1/chat/completions, so nothing else changes: same URL, same key, a different Model Name. Claude Sonnet 5.5 (claude-sonnet-5-5) is the usual first pick, at $0.0033 for one 2K-in / 300-out turn today and $3.30 for 1,000 messages.

  • Claude Sonnet 5.5: claude-sonnet-5-5, $0.0033 a message, $3.30 per 1,000, 1M tokens context.
  • Claude Opus 5.5: claude-opus-5-5, $0.0083 a message, $8.34 per 1,000, 1M tokens context.
  • Claude Haiku 4.5: claude-haiku-4-5, $0.0021 a message, $2.10 per 1,000, 200K tokens context.
  • Claude Fable 5: claude-fable-5, $0.0105 a message, $10.50 per 1,000, 1M tokens context.

Claude keeps Anthropic's own content rules whichever URL it is reached through; it is the model to pick for prose and memory, not for a card a filtered model refuses. Reasoning on the newer Claude ids counts as output tokens, so give it room in max tokens. The Claude API proxy page explains what a hosted Claude relay is and how it differs from the format translators on GitHub, and Claude by API, in crypto prices every Claude id for a key paid in USDT.

05DeepSeek proxy for Janitor AI

DeepSeek is the model most Janitor proxy guides are written around, because the stock model is cheap and the uncensored builds of it answer a card without the filter. On this key the DeepSeek ids go in Model Name like any other: deepseek-v4.1-flash-1017-uncensored ($0.0026 a message), deepseek-v4.1-flash-abliterated ($0.0022 a message), deepseek-v4-pro-uncensored ($0.0025 a message) are the builds in the table above. A -thinking build reasons first and needs a larger max tokens; a plain build answers at once.

The DeepSeek on Janitor AI guide lists every DeepSeek id on the key, uncensored, abliterated and stock, each with its context window and the price of one message, and says which ones think first.

06What 1,000 messages cost

A month of Janitor is counted in messages, and a message late in a chat sends far more than 2,000 tokens: Janitor resends the card, the persona and as much history as its context setting allows every time. The same 300-token reply, 1,000 times, at three history sizes:

Tokens sent per messageGLM-5.3 Flash UncensoredUncensored AI RoleplayClaude Sonnet 5.5
4,000$0.5317$3.10$5.70
8,000$0.9528$5.50$10.50
16,000$1.80$10.30$20.10

USD for 1,000 messages at the live rates. The input side is what grows; the output price only matters for the reply itself. A model that reports cached input bills those tokens at its cached rate, which lowers a chat that resends the same history.

There is no free tier. A $10 top-up covers about 31,142 short messages on GLM-5.3 Flash Uncensored; before the first top-up the console chat runs 12 free turns a day, enough to try a model. Swipes and regenerations are requests too, billed the same way, and a request that fails is not charged. The console's Usage page lists every request with its tokens and charge, each with a receipt at GET /v1/requests/:id/receipt you can check on the verify page. The pricing page has the billing rules and the cost calculator prices any model at any message size.

07Failover: what happens when a model goes down

Most Janitor proxies are one host behind one URL, and when that host returns a 500 or a 502 the chat shows "Provider returned error" until it recovers. Here a model id can have more than one host behind it. When a host answers an error, the API hands the request to the next host for that model, at the same price unless that host is a backup route that bills at its own rate; Janitor sees one reply. A 503 means no host for that model answered, and the request is not charged.

Two settings reach further. Add :nitro to the id in Model Name to ask for the fastest route for that request, or :floor for the lowest-cost one; with :nitro a request can go to a backup route that bills at its own rate, which can be higher than the listed price. And a custom model is a model id you name on any catalog model with up to five fallback models of your own choosing, billed at whichever model answered; type that id into Model Name and the chat moves to the next model on your list when the first fails. The status page shows each model's daily canary and open incidents.

08When it does not work

Janitor prints the status code the API returned under its own heading. The short version: 401 is the key (a typo or a space around it), 402 is an empty balance or a key that reached its budget, 404 is either a Model Name that is not an exact id or a Proxy URL that stops at /v1, 429 is the key's per-minute limit after fast swipes, and an empty bubble on a thinking build is max tokens spent on thinking. To separate a key problem from a Janitor setting, list the models the key can call from any terminal:

bash
1curl https://api.relayswitch.io/v1/models -H "Authorization: Bearer sk-relay-..."

A working key returns a list whose id values are the exact strings for Model Name. The Janitor AI proxy not working page has every message Janitor prints, including "No response from bot" and "proxy not allowed", with the fix for each.

09Sources

Menu path, field names, the full-URL rule, the relay from Janitor's servers and the save-and-refresh flow are from Janitor's help center (help.janitorai.com): the TL;DR Quickstart Proxy Instructions article and its proxy error guides, read 2026-11-02 and again 2026-11-16. Prices, context windows and tags come from the RelaySwitch catalog at render. The same key runs SillyTavern, Chub AI and RisuAI; only the URL shape differs, and each guide states it.

Questions

What does proxy mean in Janitor AI?
The Proxy option under Provider in a chat's API Settings. It sends every message of the chat to a chat completions URL you supply, with your API key and a model id, instead of to Janitor's own model. The reply comes from the model you named and the API bills you per token; Janitor relays the request from its own servers and charges nothing for it.
What is the best Janitor AI proxy?
One that publishes its prices per token, gives you your own key with a budget, runs the models you want, and retries a failed host instead of showing you the error. Here one 2K-in / 300-out turn on GLM-5.3 Flash Uncensored costs $0.00032 at today's rate, every reply carries a signed receipt, a key can be narrowed to a model list with a USD budget, and a model with more than one host fails over on its own.
Is there a Claude proxy for Janitor AI?
Yes: the same Proxy URL (https://api.relayswitch.io/v1/chat/completions) and key with a Claude id in Model Name. Claude Sonnet 5.5 is claude-sonnet-5-5, $0.0033 for one 2K-in / 300-out turn today. Every Claude id in the catalog answers the OpenAI format Janitor sends; Claude keeps Anthropic's own content rules.
Is there a DeepSeek proxy for Janitor AI?
Yes. Type a DeepSeek id into Model Name with the same URL and key; DeepSeek V4.1 Flash 1017 Uncensored (deepseek-v4.1-flash-1017-uncensored) costs $0.0026 a message today. The uncensored and abliterated DeepSeek builds answer a card without the filter; the stock ids keep it. The DeepSeek on Janitor AI guide lists every id with its price.
What is the proxy URL for Janitor AI?
https://api.relayswitch.io/v1/chat/completions, the full path. It is fixed; nothing is generated per account, only the key is yours. Paste it with no trailing slash and no extra segment. A URL that stops at /v1 works in SillyTavern and Chub, which add the rest, and fails in Janitor, which does not.
Is a Janitor AI proxy free?
Not here. Keys are prepaid from $10 by card or crypto, with no subscription needed and no KYC, and every reply bills its tokens from that balance; that first top-up covers about 31,142 short messages on GLM-5.3 Flash Uncensored. Before the first top-up the console chat runs 12 free turns a day. A free proxy is someone else's key with a queue in front of it, and it ends when the key does.
How much do 1,000 messages cost on Janitor AI through a proxy?
From $0.3211 on GLM-5.3 Flash Uncensored to $10.50 on Claude Fable 5 for 1,000 messages of 2,000 tokens in and 300 out, at today's rates. A long chat sends more history per message, so the table on this page prices the same 1,000 replies at 4K, 8K and 16K of history.
Which models work best for roleplay on Janitor AI?
The uncensored and abliterated builds for a card the filtered models refuse: they answer the scene as written with no jailbreak prompt. Claude for prose and memory across a long chat, at a higher price and with its own rules. The ranked list on the best models for JanitorAI page shows what other users of this API actually call most, and the table above prices each.
Does a proxy work on every Janitor AI character?
No. The character's creator decides whether proxies are allowed when editing the character, and the proxy panel cannot override it. If a chat says proxies are not allowed, pick another character or allow proxies on one you made yourself.
What happens when the model behind my proxy goes down?
When a host serving the model answers an error, the API hands the request to the next host for that model, at the same price unless that host is a backup route that bills at its own rate, and Janitor sees one reply. A 503 means no host for that model answered; the request is not charged. Switch Model Name to another id, or type a custom model id with your own fallback list, and the chat continues.
Can I use the same proxy for SillyTavern or Chub?
Yes, the same key and models. SillyTavern, Chub and Risu take the base URL https://api.relayswitch.io/v1 and add /chat/completions themselves; Janitor takes the full https://api.relayswitch.io/v1/chat/completions. The SillyTavern and Chub guides on this site have each app's field names.

One Proxy URL, every model

Prepaid from $10 by card or crypto. No KYC, no subscription needed.

Join on Telegram