Skip to main content

​
Choose an integration path

Public documentation preview — incomplete. The package and optional AI SDK peers remain private and unpublished. These integration paths are guidance, not a qualified public release or a hosted service.
Use the smallest path that meets your task. All execution paths require trusted host policy; results can be denied or uncertain.
  1. You know the exact tool and arguments. Register a local tool and call it, or obtain a current bound descriptor from discovery/filtering and call it with its expected snapshot. No inference is necessary.
  2. You need relevant tool candidates but no side effects. Filter tools. The default lexical ranker is model-free; named ranking uses an explicitly configured .ranking adapter.
  3. You want one proposed action. Scope a bounded execute to the intended target or compatible model group. The selected routing adapter proposes arguments; validation, policy and approval still gate the one possible call.
  4. Your host already uses the AI SDK. Build a governed ToolSet. The host chooses the generateText model; each emitted tool call goes through router.call with fresh trusted context. The host controls step and repeat-call budgets.
The common models → servers → policies configuration is optional for model-free local calls. Bindings choose routing, not the ranker used by filtering and not the AI SDK host’s model. Neither bindings nor filtering authorize effects. This SDK does not supply an implicit provider, autonomous multi-step planner, automatic specialist picker, browser runtime or hosted gateway. Do not turn needs_input, pending_approval or unknown side effects into automatic retries.