KoboldAI Lite on a custom OpenAI-compatible endpoint.

Quick answer: the values Lite needs.

Open the AI connection panel (the AI button in Lite's top menu), switch from the default AI Horde to the custom endpoint group, and pick the OpenAI-compatible option. Exact labels shift slightly between builds, but the fields are constant: an API URL, a key, and a model picked after connecting. Two Lite-specific details do most of the damage when this fails. First, the version checkbox: Lite offers an "Add Ver. Num" style toggle that appends /v1 to the URL for you, so entering https://api.apisrouter.com with the box ticked and https://api.apisrouter.com/v1 with it unticked both resolve correctly, while doubling the /v1 does not. Second, persistence: Lite has historically reset custom endpoint URLs on reconnect, so keep the URL handy rather than assuming the field survives every session.

API URL:  https://api.apisrouter.com/v1   (or base URL + "Add Ver. Num" ticked)
API key:  sk-...                            (from APIsRouter)
Model:    deepseek-v4-flash                 (or any catalog id after connecting)

What Lite is, and how it reaches a hosted model.

KoboldAI Lite is the zero-install web UI of the Kobold ecosystem: a single-page app for story writing, adventure mode, and character chat that runs entirely in your browser. It ships in two forms that share the same settings: the hosted instance at lite.koboldai.net, and the bundled copy KoboldCpp serves on localhost when you run models locally. The project README describes it as capable of connecting to custom endpoints including OpenAI-shaped and Claude-shaped APIs, alongside local and remote Kobold instances and the crowdsourced AI Horde. Because Lite is a browser app with no server of its own, your API requests go directly from the browser to whatever endpoint you configure. That design is why the custom-endpoint route works at all, and also why the CORS section below exists: the endpoint has to answer browser-origin requests, and when one does not, Lite offers a proxy toggle with tradeoffs worth understanding before flipping it. For people arriving from AI Horde: the appeal of a metered endpoint over the volunteer cluster is predictability. Horde is genuinely free with a queue; a funded endpoint answers immediately with the exact model you asked for, and at value-family rates an evening of play costs a small fraction of a cent per message.

Setup, field by field.

If the model list stays empty after connecting, the usual causes rank: a doubled or missing /v1 (check the version checkbox first), a key pasted with whitespace, or a browser extension blocking the request. The same three values working in a curl command is the fastest way to prove the endpoint side is fine and localize the problem to the browser.

• Open Lite (hosted or bundled) and click the AI button in the top menu to open the connection panel.
• Switch the provider selection from AI Horde to the custom endpoint group, and pick the OpenAI-compatible option (labels vary slightly by build; a Claude-shaped option sits alongside it).
• Enter the URL: https://api.apisrouter.com/v1, minding the version checkbox described above so /v1 appears exactly once.
• Paste your sk-... key into the key field.
• Connect, then pick a model from the list Lite fetches, or type the exact catalog id if your build asks for a manual model name.
• Send one short message in a fresh story to confirm the round trip before loading a long session.

The Claude endpoint option, and when to use it.

Lite's custom endpoint group also includes a Claude-shaped option for Anthropic-dialect APIs. APIsRouter serves that dialect too, at /v1/messages, so claude ids are reachable through either door. In practice the OpenAI-compatible route is the simpler default even for Claude models, because one configuration then serves every family: claude-sonnet-4-6 for high-end prose, deepseek-v4-flash for volume play, glm-5.2 and kimi-k2.6 for rotation, all as model-field changes rather than reconfigurations. Where the Claude-shaped option earns its place is if you keep presets tuned specifically for Anthropic's request format or migrate settings from another Anthropic-dialect tool. Functionally, both routes end at the same models through this gateway; pick one and stay consistent so your saved settings remain portable.

Samplers and modes: what actually applies over a hosted endpoint.

Lite exposes the full local-model sampler zoo in its settings, and most of it does not travel. On OpenAI-compatible endpoints, Lite's own documentation notes that only temperature, top-p, and repetition-penalty style controls apply; exotic samplers like Min-P, Top-A, or TFS are local-inference features, and while Lite can attach them as extra fields, hosted endpoints generally ignore or reject non-standard parameters. Set temperature around 0.8 to 0.9 for prose variety, leave top-p near 0.95, and keep repetition penalties low; that is the whole tunable surface, and it is enough. Lite's writing modes all work over the custom endpoint, with one guidance: instruct mode maps most naturally onto hosted chat models, since it produces the structured role-based exchange those models are trained on. Classic story co-writing works too; expect models to behave more like conversational partners than raw text continuers, because that is what a chat completions endpoint is. Context budgeting follows the same rule as every roleplay frontend: Lite resends the visible story context each turn, so set the context size deliberately (a 16K to 32K working window covers long sessions) rather than maxing it because the model's window allows it. The context-length guide linked below does this math across models.

Pay-as-you-go · transparent per-model pricing

Selected models are priced below official list prices. Exact input, output, cache, and per-request prices are shown for each model.

CORS, the proxy toggle, and keeping your key sane.

Because Lite calls the endpoint straight from your browser, a connection failure with correct values is usually CORS: the endpoint declining browser-origin requests. Lite ships a "Use CORS Proxy" toggle for exactly this case, and it deserves an honest warning: routing through any third-party proxy places that proxy inside your traffic, key and story content included. Try a direct connection first, confirm the endpoint works via curl if in doubt, and treat the proxy toggle as a last resort rather than a default. Running the bundled Lite from a local KoboldCpp is another clean workaround, since the same custom-endpoint settings exist there. Key hygiene matters a little extra on a browser app: the key sits in your browser's local settings, so avoid pasting it into shared or public machines, and if it ever leaks, revoke and reissue rather than hoping. Keys here are free to create, so a dedicated key for Lite keeps the usage log readable and the blast radius small. One content note, stated plainly: a custom endpoint changes where requests go, not what models permit. Upstream model policies and the terms of whatever Lite instance you use still apply; the policies-compared page linked below covers the landscape factually, model family by model family.

FAQ