Model providers

Each subagent in a clan can run on a different model. Same agent.py, same deepagents runtime; the [model] block in each subagent's deepagents.toml selects the provider.

How it's configured

Each subagent's deepagents.toml declares its model:

[model]
provider = "anthropic"   # which provider
name = "claude-sonnet-4-6"  # which model

The orchestrator's deepagents.toml at the repo root sets the default; each subagent overrides as needed.

Supported providers

Lazy loading

Provider packages are imported only when a subagent requests that provider. Missing keys do not block agent startup — the provider is simply unavailable until the key is set.

If a subagent's config asks for a provider whose key is not set, the agent build logs a warning and the subagent inherits the orchestrator's model.

Adding a new provider

Two steps:

1. Add the LangChain integration package to platform/requirements.txt (e.g. langchain-fireworks).
2. Extend _RESOLVERS in platform/api/model_registry.py with a new entry.

def _fireworks(name: str, **kwargs) -> BaseChatModel:
    from langchain_fireworks import ChatFireworks
    return ChatFireworks(model=name, **kwargs)

_RESOLVERS = {
    ...,
    "fireworks": _fireworks,
}

Subagents can now use provider = "fireworks".

Bring your own model endpoint

The ollama and openai providers accept a custom base_url, so a subagent can target an Ollama instance running on a laptop tunnel, Docker Cloud, a VPS, or a GPU host. Same code path on self-host and humancensys.com.

OLLAMA_BASE_URL=https://your-endpoint.example.com
OLLAMA_AUTH_HEADER=Bearer your-long-random-secret

Setup walkthroughs (laptop tunnel, Docker Cloud, AWS/GPU): BYO personal Ollama.

Two-mode notes

Connection to Pillar 1

The agent-builder UI (Step 3 of the agent-builder flow) groups providers as:

• BYO subscription: Anthropic, OpenAI, Google
• Open-weight hosted: Hugging Face, Together, Groq
• Local: Ollama