OpenRouter is the free Janitor AI proxy most people end up with, and for a simple reason: one account and one key put every free model on OpenRouter behind Janitor's proxy field at once. Janitor AI's own beginner guide calls a proxy "simply a bridge", the service that connects JanitorAI to other models, and OpenRouter is the widest bridge there is. This page gives you the exact values, the current list of free models, the limits in OpenRouter's own numbers, and the two settings that produce most of the errors.
One thing to clear up first, because it costs people evenings. The proxy here is a language-model endpoint, not a network proxy that changes your IP address. We run a proxy network of the other kind, which is how we see the confusion daily and why we have no reason to sell you the wrong thing. Where a network proxy fits, and where it does not, is covered at the end.
The OpenRouter proxy URL and the other values to paste
Open a chat, tap the menu (the three lines), choose API Settings, then Proxy, then Add Configuration. Janitor's guide names five fields, and only four matter for OpenRouter.
| Field | What goes in | Notes |
|---|---|---|
| Config Name | anything, for example OpenRouter free | a label for you |
| Proxy URL | https://openrouter.ai/api/v1/chat/completions | the full path; Janitor's error guide says it must be exactly this |
| API Key | your key from openrouter.ai, begins sk-or- | create it on the Keys page; treat it like a password |
| Model Name | a current :free id, for example nvidia/nemotron-3-ultra-550b-a55b:free | case sensitive, suffix included |
| Custom Prompt | optional | a system instruction; leave empty at first |
Janitor wants the full path in the Proxy URL field. A client that asks for a base URL instead, such as the OpenAI SDKs in the OpenRouter quickstart, takes https://openrouter.ai/api/v1 and adds /chat/completions itself.
Save, then send a short test message. If it answers, you are done. If it does not, the error section below covers every case Janitor's own guide lists.
The free models, as of 28 September 2026
OpenRouter tags free models with a :free suffix, and the list changes as providers add and retire them. We pulled OpenRouter's public models API on 28 September 2026 and found 16 ids carrying the tag, two fewer than on 24 August. The larger ones, by context window:
thinkingmachines/inkling:freeandthinkingmachines/inkling-small:free(1,048,576 tokens)nvidia/nemotron-3.5-lightning:free(1,000,000 tokens)nvidia/nemotron-3-ultra-550b-a55b:free(1,000,000 tokens)dots-studio/dots-3-note-preview:free(512,000)qwen/qwen3.8-27b:freeandinclusionai/ling-3.0-flash-sante:free(262,144)poolside/laguna-s-2.1:freeandpoolside/laguna-xs-2.1:free(262,144)google/gemma-4-31b-it:freeandgoogle/gemma-4-26b-a4b-it:free(262,144)nvidia/nemotron-3-super-120b-a12b:free(262,144)cohere/north-mini-code:freeandnvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free(256,000)
The other two are a content-safety model and a 2.6-billion-parameter LiquidAI model. Two ids from the August list, z-ai/glm-5.2:free and nvidia/nemotron-3-nano-30b-a3b:free, are gone. DeepSeek is not on the list: its :free slugs were retired in 2026, which is why our DeepSeek guide now points at DeepSeek's own cheap API instead.
If an id above stops working, open OpenRouter's models page, filter for :free, and copy the current spelling. The URL never changes; the model name does. Janitor's error guide gives a live example of the pattern: a retired deepseek-chat-v3:free had to become deepseek-chat-v3-0324:free, and that one has since gone too.
A long context window matters more on Janitor AI than in most chat apps, because the bot's personality, your persona, the chat memory and the advanced prompt all ride along in every request. Janitor's guide lists exactly those as the reason a 400 error appears: too many tokens in one message.
The limits, in OpenRouter's numbers
Free is real, and it is metered. OpenRouter's published limits for :free models: 20 requests a minute; 50 requests a day if you have bought fewer than 10 credits in the account's lifetime; 1,000 requests a day once you have bought 10 credits, which is about $10, once. The credit is not consumed by free models. It sits in the account, and you can spend it on paid models later or not at all.
Janitor's own beginner guide says the same thing in its words: 50 free messages every day, and 1,000 a day after adding $10 in credits that do not need to be spent. For a long roleplay evening the difference between 50 and 1,000 is the difference between a cap you hit before dinner and one you never see.
Two more rules from the same page decide what does not help. A negative balance blocks free models too: OpenRouter says you may see 402 errors, including for free models, until the balance is above zero. And the sentence that settles every "just make another account" thread: making additional accounts or API keys will not affect your rate limits, because OpenRouter governs capacity globally. A second account, a second key or a different IP address all leave the cap where it was.
The privacy setting behind most 404s
Many free models on OpenRouter are served by providers that require permission to use your prompts, and OpenRouter will not route to them unless you allow it. With the setting off, Janitor shows a 404 that usually reads as no endpoints found. Janitor's error guide gives the fix in one line: enable Model Training under Privacy Settings on OpenRouter.
Read that trade honestly before you flip it. Roleplay logs are personal, and the toggle means the providers behind the free models may train on them. If that is not acceptable, use a paid model on OpenRouter with the setting off, a local model, or DeepSeek's own cheap API, where the terms are DeepSeek's rather than a rotating cast of free providers. We would rather you make that choice on purpose than discover it in a settings menu.
The errors, in Janitor's own list
Janitor AI publishes an OpenRouter error guide, and it covers the whole set. Short version:
- 400. Too many tokens in one request: bot personality plus your message plus chat memory plus advanced prompt. Shorten, trim memory, or pick a bigger-context model.
- 402. Insufficient balance. Use a
:freemodel, or bring the balance above zero. - 404, three ways. Provider settings restricted too far (unblock Ignored Providers, clear Allowed Providers, save); a retired model id (take the current one); privacy blocking the free model (enable Model Training). The dedicated page is Janitor AI no endpoints found.
- 408. Servers slow or overloaded; wait and retry.
- 429, two ways. The daily cap, or a specific model overloaded. Wait, change model, or add the $10 once. The dedicated page is Janitor AI error 429.
- 502. The provider behind the model broke; check the model's uptime tab on OpenRouter, switch or wait.
- Unknown response. Usually the message limit; Janitor says to wait for the reset at 12:00 UTC.
- Network error. The URL is wrong or you did not refresh after saving. It must be exactly the Proxy URL in the table above, with no extra slashes.
The shared-proxy trap
Discord and Reddit are full of ready-made OpenRouter "proxies": somebody else's URL and key, no sign-up needed. Do not use them. You are spending a stranger's credit, which can vanish mid-scene, and the stranger's server sees every message and reply in plain text on its way through. Some exist for exactly that. OpenRouter's own key page takes two minutes, and the whole point of this guide is that the free tier is real, so there is nothing to gain from a borrowed one. Our roundup of the best free Janitor AI proxies has the longer version of the warning, and it applies to every backend, not just this one.
Where a network proxy fits
A network proxy changes the address your traffic leaves from, and nothing on this page needs one. The cap, the privacy setting and a retired model id all live in your OpenRouter account. Janitor AI has one rule that follows the address. Users connecting from IP addresses in Australia, Brazil or the UK must pass a one-time age check, and Janitor says "Detection is IP-based only." Before that check, it blocked UK addresses from 24 July 2025 until 15 June 2026. If your country, school or office network blocks OpenRouter or Janitor AI, we give no steps around that.
Where our lines do fit is a network that needs a known exit, such as a company network. For that, a static residential or ISP address in the country where you are costs from $0.44 per GB pay as you go. For everything else on this page, the fix is above: the exact URL, a current :free id, the privacy setting, and $10 once if you chat a lot.
Sources
- Janitor AI Help Center, The Absolute Beginner's Guide to Using a Proxy with Janitor, Part One and Part Two, read in their Web Archive copies of 31 July and 1 September 2026 (Janitor removed both guides by 28 September 2026): what a proxy is, the menu path, the field names, the OpenRouter URL, 50 free messages a day and 1,000 with $10 in credits.
- Janitor AI Help Center, OpenRouter Error Guide: the 400, 402, 404, 408, 429, 502, unknown-response and network-error explanations quoted above.
- OpenRouter, API rate limits: 20 requests a minute on free models, 50 a day under 10 credits, 1,000 a day at 10 or more, 402 on a negative balance, and the note that additional accounts or keys do not change limits.
- OpenRouter, public models list: the 16 ids tagged :free and their context lengths, read 28 September 2026; Quickstart: the base URL for OpenAI SDK clients.
- Janitor AI, Tough news for our UK users (July 2025) and An update for users in Brazil, Australia, and the UK (15 June 2026): the UK block and the switch to age verification; Janitor AI Help Center, Age Verification for Australia, Brazil, and the UK: detection by IP address.
