Skip to content

Compatible AI Providers

Ontheia supports any OpenAI-compatible inference server — local or remote. An endpoint implementing /v1/chat/completions is enough. If it also offers /v1/responses, individual models can be routed through the Responses API (chat_api: "responses") — required to combine reasoning with tool calling on current reasoning models. Anthropic is addressed through its own API protocol, not through OpenAI compatibility. Google Gemini, by contrast, goes through Google’s OpenAI-compatible layer (https://generativelanguage.googleapis.com/v1beta/openai/) — not through generateContent and not through the new Interactions API.

ProviderTypeLink
OllamaLocal, self-hostedollama.com
llama.cppLocal, self-hostedgithub.com/ggml-org/llama.cpp
LM StudioLocal, desktop applmstudio.ai
JanLocal, desktop appjan.ai
vLLMSelf-hosted, productiondocs.vllm.ai
ProviderProtocolLink
Anthropic ClaudeAnthropic APIanthropic.com
OpenAIOpenAI API (Chat Completions + Responses)openai.com
Google GeminiOpenAI-compatible (Google compat layer)ai.google.dev
xAI / GrokOpenAI-compatible (+ Responses)x.ai
GroqOpenAI-compatiblegroq.com
DeepSeekOpenAI-compatibledeepseek.com
OpenRouterOpenAI-compatibleopenrouter.ai
Azure OpenAIOpenAI-compatibleazure.microsoft.com
Mistral AIOpenAI-compatible ⚠mistral.ai
Ollama CloudOpenAI-compatibleollama.com

⚠ Mistral AI — limited tool compatibility: Chat and completions work reliably. Tool Calling is not fully OpenAI-compatible — parallel tool calls and some formatting details may cause errors. Recommended: mistral-large or mistral-small; other models may not support tool calling at all.

For local OpenAI-compatible providers:

FieldValue
Base URLe.g. http://localhost:11434/v1 (Ollama) or http://localhost:8080/v1 (llama.cpp)
API KeyAny value (ignored by local providers)
ModelModel name as shown in the respective app

Ollama is the easiest starting point for local models — single-command install, automatic GPU detection, large model library. Besides the local instance (http://localhost:11434/v1), Ollama also runs a hosted cloud service at https://ollama.com/v1 (e.g. for kimi-k3:cloud). Ontheia auto-detects both endpoints as OpenAI-compatible — the local one through the private-host check, the cloud endpoint through the ollama.com host. Tool calling and streaming work on both without a manual flag.

llama.cpp is the inference engine behind Ollama and many other tools. Choose it directly when you need full control over quantization and server parameters.

LM Studio and Jan provide a graphical interface for model management and expose a local API server. Ideal for users who want to run Ontheia as an agent layer on top of a desktop app.

vLLM is optimized for production deployments with high concurrency. Recommended when running Ontheia for multiple teams in an enterprise environment.

Google Gemini is wired up through Google’s OpenAI compatibility layer. Two quirks follow from that: Gemini accepts reasoning_effort together with function tools (which OpenAI rejects on the same kind of endpoint), and the layer reports reasoning tokens in no field of their own — they only show up inside total_tokens. Ontheia recovers them from there and counts them as output tokens so the cost display is not too low. Google’s new Interactions API (generally available since June 2026) is not used today; the compatibility layer is not deprecated.

OpenRouter is an aggregator — a single API key gives access to models from OpenAI, Anthropic, Google, Meta, and many other providers.