@cendor/core

Wrap your LLM client once and capture exact token counts and cost on every call — the shared foundation the other Cendor tools build on. The TypeScript port of cendor.core: shared types, an event bus, decimal-safe Money, an offline price table, provider-aware token counting, and instrument(). Every other @cendor/* package cooperates through this.

npm i @cendor/core
# provider SDKs are optional peers — install the one(s) you use:
npm i openai @anthropic-ai/sdk

Killer example — wrap once, get cost + tokens on every call

import OpenAI from 'openai';
import { instrument, bus, LLMCall } from '@cendor/core';

const client = instrument(new OpenAI());

bus.subscribe((e) => {
  if (e instanceof LLMCall) console.log(e.model, e.usage?.totalTokens, e.cost?.toString());
});

await client.chat.completions.create({
  model: 'gpt-4o',
  messages: [{ role: 'user', content: 'hello' }],
}); // → gpt-4o 152 0.000... USD

Streaming, the Responses API, Anthropic, interceptors (record/replay), and Reroute (model downgrade / message redaction) all flow through the same instrument().

Surface

Parity & conformance

Field names map snake_case (Python) → camelCase (TS); type and error names are identical (UnknownModelError in both). Cost math, the model table, Money semantics, and token counts are verified against golden vectors generated from the Python reference — see fixtures/. Money is compared by exact decimal value.