Point KoboldAI Lite's custom API at uncensored models, no GPU needed

Connect lite.koboldai.net to this API in one popup: pick OpenAI Compatible API, paste the URL and key, tick Chat-Completions, leave the CORS proxy off.

Steps checked October 24, 2026 · prices as of November 16, 2026

Lite, the KoboldAI front end that runs in a browser tab, can write with uncensored builds hosted on this API. The connection is one dialog: the OpenAI Compatible API provider, a base URL, a key and a model id, with Chat-Completions on and the CORS proxy off so your key travels straight from the browser to the API. No GPU, no download, no KoboldCpp. Every label here is from Lite's source as of 24 October 2026.

New to models with the refusals trained out? Read AI with no restrictions first, then come back and connect one.

What you need

  • A browser with lite.koboldai.net open. Nothing to install.
  • A key from the console. It begins with sk-unr-, appears once, and spends a balance you prepay from $10.
  • A model id from the model list.

No account at all? /chat runs the same builds in the browser with three free turns a day.

Six values, one popup

Lite keeps the whole connection in a single dialog. Fill it with these.

Provider
OpenAI Compatible API
Pick it from the Select your AI provider dropdown
OpenAI API URL
https://api.unrestrictable.ai/v1
Leave Add Ver. Num ticked
API Key (Required)
sk-unr-live-…
Generated in the console, displayed once
Model
cydonia-24b-v4.3
Choose after Fetch, or enter it under Custom
Chat-Completions API
ticked
Use CORS Proxy
unticked

Step by step: Lite to the API

  1. Make a key in the console. Lite accepts a key with an empty balance; the first generation then fails with 402.
  2. At lite.koboldai.net, click AI at the top. The Select your AI provider dialog appears.
  3. Choose OpenAI Compatible API from the dropdown.
  4. Put https://api.unrestrictable.ai/v1 in the URL box (placeholder text: OpenAI API URL) and leave Add Ver. Num ticked. Lite's hint beside the box says OpenAI-compatible endpoints end in /v1 and to untick it when yours does not. Ours does.
  5. Paste the key into API Key (Required).
  6. Press Fetch to pull the ids your key can use and pick one, such as cydonia-24b-v4.3. Or press Custom and type an id exactly as it appears on the model list.
  7. Tick Chat-Completions API so Lite sends each turn as chat messages to /chat/completions, which every model here serves. Unticked, Lite falls back to the legacy text endpoint, which is outside this guide.
  8. Keep Use CORS Proxy off. Lite's own wording is that keys go to the API directly unless the proxy is used. Off means browser to API, nothing in between.
  9. Press Ok. You are connected.

Favorites, in the same dialog, stores the URL and key in the browser so the next visit skips the paste. Do not use it on a computer other people share.

What Lite is, and what it is not

KoboldAI Lite is a single web page for stories, text adventures and character chat. You can use it at lite.koboldai.net, and it is the same interface KoboldCpp ships for local models. The code is public at LostRuins/lite.koboldai.net on GitHub.

Lite contains no model. Every generation goes to a backend you choose: KoboldCpp on your own hardware, the AI Horde, or a hosted endpoint. Picking OpenAI Compatible API makes this API that backend, which is why the GPU, the download and KoboldCpp all drop out. The field names in this guide are from Lite's source as it stood on 24 October 2026.

Usage Mode, Max Output, Context Size

Three fields in the Settings panel shape the session and the bill.

  • Usage Mode: Story, Adventure, Chat or Instruct. Choose by what you are writing; the API connection does not change.
  • Max Output: the cap on a single reply. 512 tokens is our starting point for chat and adventure.
  • Context Size: how much story Lite resends each turn. Hold it below the model's window minus Max Output.

The dialog also tells you only Temperature, Top-P and Repetition Penalty reach this provider. Non-Standard Fields forwards the others (Min-P, Top-K and so on), with Lite's own warning that many endpoints reject them. Leave it off unless a sampler you need is behind it.

Memory and World Info live in the Context panel. Memory goes out with every turn; a World Info entry goes out when one of its keys shows up in the text. Both eat Context Size and both are billed as input each time they are sent. Across the 150 turns of the example below, a 500-token Memory is 500 input tokens 150 times over, so trim it to what the story cannot do without.

Adjust Temperature before anything else: down when the build wanders off the plot, up when it repeats itself. A Repetition Penalty slightly above 1 steadies a small model that gets stuck on a phrase late in a long adventure.

Four roleplay builds and the Context Size for each

Each build here is tagged roleplay in the catalog and answers on the OpenAI wire format. Prices and windows come from the live catalog at render time.

Uncensored models and their prices per million tokens
Wire formats
MythoMax L2 13Bmythomax-l2-13bRoleplay$0.404$0.4044Kopenai
Stheno 8B v3.2stheno-8b-v3.2Roleplay$0.7035$0.703516Kopenai
Cydonia 24B v4.3cydonia-24b-v4.3Roleplay$0.484$0.60131Kopenai
Gemma 4 12B Starry Oceangemma-4-12b-starry-oceanRoleplay$0.175$0.875131Kopenai

Prices per million tokens, read from the live catalog on 2026-11-16.

Model idCatalog windowContext Size to set
mythomax-l2-13b4,096 tokens3,584
stheno-8b-v3.216,384 tokens15,872
cydonia-24b-v4.3131,072 tokens15,872
gemma-4-12b-starry-ocean131,072 tokens15,872

Windows read from the live catalog on 2026-11-16. Context Size is the window, capped at 16,384, less 512 of Max Output.

mythomax-l2-13b has only a 4,096 tokens window, so Context Size must stay below that and older story drops out fast. The 24B tune and the Gemma 4 tune hold far more history. Other Gemma tunes are on the uncensored Gemma page.

Pricing a full adventure session

Assume 150 turns in Adventure mode with replies of about 200 tokens. The first column resends about 2,000 tokens of story each turn; the second resends 8,000, which is a bigger Memory plus World Info.

Model150 turns, 2,000 tokens of context150 turns, 8,000 tokens of context
MythoMax L2 13B$0.13over the window
Stheno 8B v3.2$0.23$0.87
Cydonia 24B v4.3$0.16$0.60
Gemma 4 12B Starry Ocean$0.08$0.24

Worked out at render from the catalog prices on 2026-11-16: turns × prompt tokens at the input rate, plus turns × 200 at the output rate. "Over the window" means that much context does not fit the model.

Almost all of it is input. 150 × 2,000 comes to 300,000 input tokens, against 30,000 output tokens. Context Size is the knob that moves the total. Per-token rates for every model are on the pricing page.

Check the key from a terminal first

This call returns the model ids your key can use. Fetch in Lite makes the same request. It is free and settles the key question before you open Lite. Replace the key and run it.

curl
1curl https://api.unrestrictable.ai/v1/models \
2 -H "Authorization: Bearer sk-unr-live-…"

200 plus a list of ids: good key. 401: the key is the problem, and the body says how:

http
1HTTP 401
2{"error":{"message":"No API key was sent.","type":"authentication_error","param":null,"code":"invalid_api_key"}}
http
1HTTP 401
2{"error":{"message":"The API key is not valid.","type":"authentication_error","param":null,"code":"invalid_api_key"}}

Both bodies verbatim from the API on 24 October 2026.

Fixing Lite errors

401 (key)

"No API key was sent." is an empty key field. "The API key is not valid." is a wrong, truncated or revoked key. Make a new one in the console and paste all of it.

404 model_not_found (model id)

The id is not in the catalog, either a typo under Custom or a build that has been retired. Press Fetch again and choose from the list.

404 on each turn (Chat-Completions off)

Fetch succeeds but generation fails every time: check that Chat-Completions API is ticked. Off, Lite hits the text-completion endpoint, not /chat/completions.

402 (balance)

The prepaid balance has reached zero. Top up in the console and carry on.

CORS errors in the browser console

Lite calls the API from your tab. The API accepts requests from any origin and allows the Authorization header (verified 24 October 2026). The proxy adds nothing, so leave it off.

The daily real-model check sends every model the same prompt each day, which is how you know an id serves the model its name claims. Settings for other clients are on the docs page.

Questions

No. Lite is a free page with public source. You pay only for the backend it talks to, here per token from a prepaid balance. See per-token prices.

A browser interface for stories, text adventures and character chat, and also the UI inside KoboldCpp. It hosts no model and forwards your text to whatever backend you connect.

In your browser’s local storage. Lite autosaves the open story and settings when you leave and restores them next time. Favorited URL and key pairs live there too.

No. This API accepts browser requests from any origin. With the proxy on, your key goes through the proxy before reaching the API.

Yes. Pick OpenAI Compatible API and the model runs on this API’s side. A GPU only matters for KoboldCpp on your own machine.

No. These are abliterated or uncensored fine-tunes: the behaviour is in the weights. Nothing has to sit in Memory or World Info to make them answer. See AI with no restrictions, explained.

Every model above is on the uncensored model list with its live price per million tokens. Per-token pricing sets the whole catalog side by side, and the family directory groups the builds by base model.

Get a keyPrepaid from $10, no subscription.

New models and price changes, once a month.

New uncensored builds with their prices, price moves, and what the daily model check found.

Join on Telegram