LLM Providers¶

Soulacy supports multiple LLM providers simultaneously. Each agent picks its provider and model with llm.provider and llm.model in SOUL.yaml.

Reference¶

llm:
  default_provider: openai   # used when model has no prefix
  timeout: 120s              # per-request LLM timeout

  providers:
    openai:
      api_key: "sk-..."
      base_url: https://api.openai.com/v1   # override for proxies

    anthropic:
      api_key: "sk-ant-..."

    ollama:
      base_url: http://localhost:11434

    together:
      api_key: "..."
      base_url: https://api.together.xyz/v1

    qwen:
      api_key: "..."
      base_url: https://dashscope.aliyuncs.com/compatible-mode/v1

Selecting a model in SOUL.yaml¶

Set the provider and model explicitly:

llm:
  provider: openai
  model: gpt-4o-mini

For OpenAI-compatible routers such as OpenRouter, Together, Groq, vLLM, or LM Studio, use the configured provider plus that service's model ID:

llm:
  provider: openai
  model: moonshotai/kimi-k2

Supported providers¶

Provider	Config key	Notes
OpenAI	`openai`	GPT-4o, GPT-4o-mini, o1, o3
Anthropic	`anthropic`	Claude 3.5, Claude 3 Haiku
Ollama	`ollama`	Local models, no API key needed
Together AI	`together`	Open-source models at scale
Qwen (Alibaba)	`qwen`	Qwen-Max, Qwen-Plus, Qwen-Turbo

Custom / compatible endpoints¶

Any OpenAI-compatible API (vLLM, LM Studio, Groq, etc.) works via the base_url override:

llm:
  providers:
    openai:
      api_key: "not-needed"
      base_url: http://localhost:1234/v1   # LM Studio

Timeouts¶

llm:
  timeout: 180s   # increase for slow models or long context

Per-agent output caps are configured with llm.max_tokens; memory/context injection is controlled by the agent's memory block. See the SOUL.yaml reference.