weftai-providers

Present a weftai runtime as a tool on a model host. weftai never calls a model itself: each adapter turns one bound plan tool into the tool definition a host documents, and turns the host's tool calls back into plans. Your application keeps its own SDK or HTTP client.

pip install weftai-providers
pip install "weftai-providers[anthropic]"   # adds the Anthropic SDK, for its tool runner

from weftai.providers.openai import PRESETS, openai_tools

tools = openai_tools(
    runtime,
    {
        **PRESETS["qwenIntl"],
        "model": "qwen-plus",
        "ctx": ctx,
        "tools": [{"name": "query_tickets", "include": lambda op: op.effects == "read"}],
    },
)

Families

Each family is its own module, so an application imports only the ones it uses.

PRESETS holds the base URL and settings of hosted, local and regional OpenAI-compatible hosts, so a preset plus a model name is a complete configuration. The capability catalog checks known model and API combinations, and names an explicit override when a model is unknown.

Decisions over HTTP

weftai's decisions interface asks a Decider short, typed questions (a choice, a score, a yes or no) and gets back answers with probabilities. LayaDecider is one implementation, for a decision service reached over HTTP: you supply its endpoint and an httpx.AsyncClient, which you own. Failures and timeouts abstain rather than guess, and what a host does with an answer, including any permission, stays the host's decision.

License

MIT