Janitor AI proxy: what it is, how tokens work, and the risks

In Janitor AI, a "proxy" is the connection to an AI model that does not belong to Janitor. Janitor's glossary defines it as "the middleman between your chat and the model provider". By default, chats run on JLLM, Janitor's own model. A proxy lets you use a model from another company instead, usually by entering that provider's address and your own API key.

This page explains the options using Janitor's help center, checked on October 23, 2026, and the risks that come with each. It does not cover ways around any provider's content rules. For Janitor's overall privacy picture, see Is Janitor AI safe?.

Option	Who runs the model	What you need	Cost
JLLM	Janitor	Nothing; it is the default	Free; janitor+ has JLLM+
janitor+ Router	Third-party models, accessed by Janitor	A janitor+ subscription, web only	Per token, from a prepaid wallet
Own API configuration ("proxy")	An outside provider you choose	An account and API key with that provider	Whatever the provider charges

JLLM is described in the glossary as "free, built-in, and ready to go out of the box".

janitor+ Router, per its help article (last updated July 31, 2026), lets subscribers use third-party models "without setting up an external proxy or pasting in API keys". Usage is billed per token at rates shown before you chat, subscribers get 10% off some popular models and a one-time $1 launch credit that expires a year after claiming, and Router credits are non-refundable except where law requires. Janitor still recommends JLLM+ for everyday roleplay.

Own API configuration is what most people mean by "Janitor AI proxy". In the chat's API settings you add a configuration with a model name, the provider's endpoint URL and your API key. Janitor's quickstart walks through this with OpenRouter, a model marketplace, and a DeepSeek model as the example. Free tiers and limits are set by the provider, not by Janitor, and change over time, so check them on the provider's own pricing page.

Tokens are the small chunks of text a model reads and writes. Janitor's article Tokens: your AI's memory budget explains the two things they control:

• Memory. A model can only hold a certain number of tokens at once. The article says JLLM's budget is often between 8,000 and 9,000 tokens. Once a chat passes the limit, the oldest parts drop out, which is why a bot "forgets".
• Cost. With Router or a paid outside provider, you pay per token, for both your messages and the replies. Long character definitions and long chats use more tokens on every message.

The glossary gives a rule of thumb: roughly 1,000 tokens equals 750 words. Each model counts slightly differently, so online counters are estimates. The help center says the character creation page shows the accurate JLLM token count for a character.

"Janitor AI prompts" in the help center mean the instructions that shape a model's behavior: the character definition and an optional custom prompt. They count toward the token budget too.

Janitor's own quickstart tells users to save the key "like a password". Anyone who has it can spend your balance with the provider. To limit the damage:

• Set a spending cap or prepaid balance at the provider.
• Never paste the key into a bot description, screenshot, Discord message or a "free proxy" form.
• Create a separate key just for Janitor so you can revoke it without affecting anything else.
• Revoke the key when you stop using it. Deleting your Janitor account does not cancel it.

Be careful with proxy URLs shared by strangers. A proxy endpoint receives every message and, often, your key. If you would not give that person your chats and your password, do not use their endpoint.

With JLLM, your chats go to Janitor. With a proxy, every message in that chat also goes to the provider you connected, and sometimes on to the company that runs the model behind it.

OpenRouter's data handling page is a useful example of what to look for. It says it works with providers to prevent training on prompts "wherever possible", "but there are exceptions", and that providers "have their own data retention policies". It also lets account holders restrict routing to providers that meet specific data policies. Read the privacy policy of each company in the chain before sending anything personal, and keep real names, addresses and photos of real people out of chats.

Connecting another model does not switch off any rules. Janitor's content guidelines still apply to your bots and chats, including its zero-tolerance policy on minor-coded characters. The provider's own terms apply as well: OpenRouter has its terms of service, and each model company has a usage policy.

Prompts or setups built to get around a provider's safety rules can break those terms and lead to a banned key or account at the provider, with any prepaid balance at stake. We do not publish them. If you want adult roleplay, the straightforward route is a service whose terms allow it for verified adults.

Proxy errors are often not Janitor's fault. The help center's rate limit and network error article says these errors "often" come from proxy providers when their servers are down or overloaded, and suggests switching to JLLM to test. Janitor also publishes provider-specific guides, such as the OpenRouter error guide. If JLLM also fails, check the official status page; our guide Is Janitor AI down? explains how.

If managing keys, balances and several privacy policies sounds like too much, there are two simpler options: stay on JLLM, or use the Router inside janitor+ so there is no key to leak. Another option is a service that hosts its own models and publishes one price list. Janitor AI alternatives compares several, and the Janitor AI review covers pricing and rules.

• Janitor help: JLLM and glossary
• Janitor help: janitor+ Router (updated July 31, 2026)
• Janitor help: proxy quickstart
• Janitor help: tokens
• Janitor help: rate limit and network errors
• Janitor help: OpenRouter error guide
• Janitor help: minor-related content
• OpenRouter: data handling and logging
• OpenRouter terms of service

Checked October 23, 2026.