Skip to main content
Model choices and provider choices never appear in your code — only logical preset names.
AgentHeaven routes all LLM calls through LiteLLM, supporting any provider it covers. Rather than hard-coding model names or API keys in your code, you define presets — logical roles like sys, chat, reason, embedder — and configure each preset to point at whichever provider and model you want via Configuration. 
from ahvn.utils.llm import LLM

llm = LLM(preset="chat")      # uses whatever model/provider you configured for "chat"
answer = llm.oracle("Hello!")  # full response as a string
This design means your application code never changes when you swap from GPT to Claude to a local Ollama model — you just update the config.

​
Core Concepts

ConceptWhat it isExample
PresetLogical rolesys, chat, reason, embedder, coder, translator, tiny, local
ProviderBackend connectionopenrouter, openai, anthropic, gemini, deepseek, ollama, lmstudio, vllm
Model aliasShort name → full IDdsv3 → deepseek/deepseek-v3.2, sonnet → claude-sonnet-4-6
BackendLiteLLM routing prefixopenai/, anthropic/, ollama/, hosted_vllm/
See Quick Setup for how to configure providers and presets.

​
LLM Features

Chat & Streaming

oracle(), stream(), message formats, and per-call parameters.

Sessions

Multi-turn conversations, history, slash commands.

Embeddings

embed(), embedding providers, vector dimensions.

Tool Use

Function calling, tool schemas, exec_tool_calls.

Advanced

Image generation, include finetuning, custom backends.

​
Further Exploration

Related: