Skip to main content

Shipped Adapters

NeoTavern ships a set of provider adapters out of the box. They live in packages/provider-sdk/src/adapters/, one file per adapter, and are registered in the core ProviderRegistry by their provider kind.

OpenAI-Compatible​

File: openaiCompatible.ts — kind openai-compatible.

Targets any server exposing the OpenAI /v1/chat/completions and /v1/models API: OpenAI itself, OpenRouter, LM Studio, llama.cpp server, Ollama with the /v1 endpoint, vLLM, and similar. It uses only the global fetch and the SDK's SSE parser; the API key is sent but never logged.

Anthropic​

File: anthropic.ts — kind anthropic.

Targets the native Anthropic Messages API. This is the one documented exception to the no-vendor-SDK rule: it uses @anthropic-ai/sdk because the API — extended thinking and beta-header support — is handled more accurately by the official SDK. It supports prompt caching and adaptive thinking and declares the assistantPrefill wire capability.

Text Completion​

File: textCompletion.ts — kind text-completion.

Targets local or self-hosted backends that expose the legacy OpenAI /v1/completions endpoint: text-generation-webui ("ooba"), koboldcpp, vLLM, Ollama, llama.cpp server, and similar. Unlike chat adapters, it consumes a serialized prompt: the prompt pipeline renders the instruct format and hands the adapter a single user message whose content is the finished prompt, and the adapter posts it to /completions. The API key is optional for local servers and never logged.

NovelAI​

File: novelai.ts — kind novelai.

Targets the NovelAI text-generation API (POST {baseUrl}/ai/generate with a Bearer key). Generation is non-streaming — a single delta plus the terminal done event, matching the unified stream contract. Model discovery is not offered by the API, so listModels returns the configured model. The adapter is marked experimental because NovelAI's parameter surface evolves; only the well-established samplers are mapped.

KoboldAI​

File: koboldai.ts — kind koboldai.

Targets the KoboldAI/Kobold server native API (POST {baseUrl}/api/v1/generate). Generation is non-streaming; the loaded model is read from /api/v1/model for discovery. Typical local installs need no API key.

AI Horde​

File: aiHorde.ts — kind ai-horde.

Targets the AI Horde (stablehorde.net), an asynchronous crowdsourced cluster. A job is submitted with /api/v2/generate/text/async, then polled via the status endpoint until done; the poll loop re-checks the caller signal and an idle deadline, so a stuck job aborts instead of polling forever. Anonymous use is allowed at lower priority; an API key is sent as the apikey header when configured.

Echo​

File: echo.ts — kind echo.

A fully offline provider used for tests, demos, and verifying the streaming pipeline without any network or API key. It streams the last user message back word-by-word. It also implements the optional speech, image, and transcription methods, which makes it a useful reference for writing an adapter that covers every modality.

Prompt Helper​

File: prompt.ts — exports promptFromMessages, a shared helper that serializes message arrays into the prompt shapes the adapters send. It is not an adapter itself.

For the exact ProviderAdapter interface all of these implement, see Adapter Contract and the generated Provider SDK reference.