@zanii/connectors

One-call proof recording for the two dominant LLM tool-calling formats — OpenAI function calls and Anthropic tool_use — plus a toolbox wrapper. It's the platform-shaped glue over the SDK's generic wrapTool, so you don't hand-map each provider's payload. Zero dependencies; the agent is duck-typed.

npm install @zanii/connectors @zanii/sdk

Usage

import { wrapToolbox, runToolCalls } from '@zanii/connectors';
import { ZaniiAgent } from '@zanii/sdk';

const agent = new ZaniiAgent({ /* … */ });

// 1. Wrap a whole toolbox — every call is recorded to tools.<name>:
const tools = wrapToolbox(agent, {
  lookupCustomer: (id: string) => db.get(id),
  sendEmail: (to: string, body: string) => mailer.send(to, body),
});
await tools.sendEmail('[email protected]', 'hi'); // recorded automatically

// 2. Or dispatch a model's tool calls and receipt each (OpenAI or Anthropic):
const outcomes = await runToolCalls(agent, message.tool_calls, async (name, args) => {
  return myDispatcher[name](args);
});
// each outcome: { name, arguments, result | error, hash }

normalizeToolCall handles both shapes (OpenAI function.arguments JSON strings are parsed; Anthropic input is passed through). runToolCalls records both successes and failures and keeps going.

Changelog

• 0.1.0 — initial release: wrapToolbox, normalizeToolCall, recordToolCall, runToolCalls.

License

Apache-2.0.