"KoboldAI Classic API" names a request shape: a raw prompt posted to /api/v1/generate, text returned in results[0].text. The first KoboldAI server exposed it; KoboldCpp still serves it; SillyTavern, Kobold Lite and many Python scripts send it to a model on the user's own GPU. Below: what the API is, why this API answers in OpenAI format instead, and the exact settings that move those same front-ends onto an uncensored build with no GPU in the picture. Every source was read on 27 October 2026.
Just connecting Kobold Lite? The Kobold Lite custom API guide gets there faster.
Where the name comes from
Front-ends use "KoboldAI Classic" for the first KoboldAI server and the HTTP interface it offered. Its README describes a browser front-end for AI-assisted writing across local and remote models, and notes that the REST API sits at /api under the URL Kobold hands you. On 27 October 2026 that README also warned that the legacy product "is currently not maintained" and sent readers to KoboldCpp.
The program faded; its API did not. KoboldCpp took over the endpoints and documents them in an API reference (version 2026.10.11 when read on 27 October 2026) under a heading for the KoboldAI United compatible core endpoints. Whenever SillyTavern, Kobold Lite or a Python script is set to "KoboldAI Classic", it posts this request shape to http://localhost:5001, the port KoboldCpp opens by default.
SillyTavern's connection docs mark the original "deprecated, abandoned", with context capped at 2k and no streaming, and give KoboldCpp its own entry as an easy API with CPU offloading for low-VRAM machines and streaming. Both sit under Text Completion. Both need your own hardware.
Endpoints and the generate body
Five core endpoints appear in the reference. Prefix each path with the server host, http://localhost:5001 for a KoboldCpp on your machine.
- POST /api/v1/generate
- Generate text with a specified prompt
- GET /api/v1/model
- Retrieve the current model string.
- GET /api/v1/info/version
- Current KoboldAI United API version
- GET /api/v1/config/max_context_length
- Retrieve the current max context length setting value that public backends see
- GET /api/v1/config/max_length
- Retrieve the current max length (amount to generate) setting value
A generate request carries one raw prompt string, no message list. The fields the reference documents: prompt (the submission), max_length (tokens to produce), max_context_length (the most tokens the model is sent), stop_sequence, the samplers temperature, top_p, top_k and min_p, then rep_pen (the base repetition penalty) with rep_pen_range, and sampler_order. The answer comes back in results[0].text.
POST http://localhost:5001/api/v1/generate{ "prompt": "The lighthouse door slammed behind her.", "max_length": 200, "max_context_length": 4096, "temperature": 0.8, "top_p": 0.9, "rep_pen": 1.1, "stop_sequence": ["\nYou:"]} 200 OK{"results": [{"text": " Rain sheeted off her oilskin..."}]}Out of the box there is no key. KoboldCpp's wiki explains that the --password launch flag adds one, after which generations need an API key; with no flag, anyone who reaches the port can generate. When someone asks for a "Kobold API key", that password is all they mean.
Why this API is not a Kobold endpoint
There is no /api/v1/generate at https://api.unrestrictable.ai/v1. Each model answers POST /v1/chat/completions in OpenAI form: a model id, a messages array, max_tokens where Kobold has max_length, stop where it has stop_sequence, and the key in an Authorization header. Here is the request above, rewritten:
POST https://api.unrestrictable.ai/v1/chat/completionsAuthorization: Bearer sk-unr-live-…{ "model": "cydonia-24b-v4.3", "messages": [ {"role": "system", "content": "You are the narrator of a storm-bound lighthouse story."}, {"role": "user", "content": "The lighthouse door slammed behind her."} ], "max_tokens": 200, "temperature": 0.8, "top_p": 0.9, "stop": ["\nYou:"]}KoboldCpp went the same way. Its reference opens with the OpenAI-compatible and Anthropic endpoints, which it recommends to beginners, serving /v1/chat/completions and /v1/completions on the usual port. Any client that talks OpenAI format to a local KoboldCpp talks it here too, once the host and key change. Temperature, top_p and stop sequences carry over; repetition penalty, sampler order and the samplers only Kobold has do not.
In practice: select Chat Completion (some front-ends label it OpenAI-compatible) instead of "KoboldAI Classic", and the request leaves in the right shape.
Settings for SillyTavern, Kobold Lite and Python
SillyTavern
API: Chat Completion. Chat Completion Source: Custom (OpenAI-compatible). Custom Endpoint: https://api.unrestrictable.ai/v1. Paste the key, press Connect, choose a model id from the list that loads. Selecting KoboldAI Classic or KoboldCpp under Text Completion never reaches this API. Every field and every error is in the SillyTavern guide.
KoboldAI Lite
Open the AI popup, pick OpenAI Compatible API, enter https://api.unrestrictable.ai/v1 as the URL and your key, tick Chat-Completions API, and keep the CORS proxy off. Field by field, that popup is covered in the Kobold Lite custom API guide.
Python
Where a script posted to /api/v1/generate, one OpenAI SDK call takes over. This helper keeps the Kobold-style signature, so nothing downstream moves:
from openai import OpenAI client = OpenAI(base_url="https://api.unrestrictable.ai/v1", api_key="sk-unr-live-…") def generate(prompt: str, max_length: int = 200) -> str: """Same signature as a Kobold /api/v1/generate helper, sent to the OpenAI endpoint.""" reply = client.chat.completions.create( model="cydonia-24b-v4.3", messages=[{"role": "user", "content": prompt}], max_tokens=max_length, temperature=0.8, ) return reply.choices[0].message.content print(generate("The lighthouse door slammed behind her."))Make the key in the console; it begins sk-unr- and is displayed one time. Balance is prepaid, starting at $10. No download, and your GPU stays idle. To try a model with no front-end at all, /chat runs the same catalog in the browser with three free turns a day and no account.
Picking a build
Most Kobold users are after story and roleplay tunes. These four carry the roleplay tag in the catalog, and MythoMax is the 13B that older Kobold guides keep recommending. Prices and context windows come from the live catalog each time the page renders.
| Wire formats | |||||
|---|---|---|---|---|---|
| MythoMax L2 13B | Roleplay | $0.404 | $0.404 | 4K | |
| Stheno 8B v3.2 | Roleplay | $0.7035 | $0.7035 | 16K | |
| Cydonia 24B v4.3 | Roleplay | $0.484 | $0.60 | 131K | |
| Magidonia 24B v4.3 | Roleplay | $0.3535 | $0.4235 | 33K |
| Model | 150 generations at 2,000 context |
|---|---|
| MythoMax L2 13B | $0.13 |
| Stheno 8B v3.2 | $0.23 |
| Cydonia 24B v4.3 | $0.16 |
| Magidonia 24B v4.3 | $0.12 |
Running KoboldCpp at home costs a graphics card and power; running here costs what the table says. The models page lists every model, the pricing page every rate, and the daily check posts one fixed prompt to each model per day to prove the id serves its labelled model.
First request: list your models
This is GET /api/v1/model in OpenAI terms: the ids your key is allowed to use. A 200 carrying a list means both the URL and the key are good.
curl https://api.unrestrictable.ai/v1/models \ -H "Authorization: Bearer sk-unr-live-…"401 points at the key: "No API key was sent." or "The API key is not valid." 404 on a chat call means a bad model id, or a Kobold path aimed at an OpenAI endpoint. 402 means the balance is empty, and nothing was charged. Settings for every other client are on the docs page.
Questions
Your own KoboldAI or KoboldCpp server address with /api appended; a local KoboldCpp answers at http://localhost:5001/api/v1/generate. A hosted uncensored model has no Kobold URL at all. Give your client the OpenAI-compatible base URL under its Chat Completion option instead. See base URL and key per client.
Only if the server owner set one: KoboldCpp's wiki describes the --password flag, after which generations require an API key. This API needs a key every time. It begins with sk-unr- and comes from the console.
To a local server: POST JSON containing prompt and max_length to /api/v1/generate, then read results[0].text. To this API: the OpenAI SDK with base_url pointed at the endpoint. The snippet above wraps that call in a Kobold-style function.
No. KoboldAI Classic is the original server, whose README called it "not maintained" on 27 October 2026. KoboldCpp is the maintained program, and it serves the same /api/v1 endpoints alongside OpenAI-compatible ones.
Yes. Choose Chat Completion with Custom (OpenAI-compatible), fill in endpoint and key; the model then runs on this API. The KoboldAI Classic and KoboldCpp options inside SillyTavern expect a server on your own computer. See the SillyTavern guide.
Neither. The Colab notebooks existed to run KoboldAI on a rented GPU. Here the model is already up, and the client needs three things: base URL, key and a model id.
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.