* Install and import `@anthropic-ai/tokenizer`. * Map each Claude model (Haiku, Sonnet, Opus) to `countTokens(text)` calls. * Verify context window limits (e.g., Opus ≈ 200 k tokens). * Add fallback to `text.length / 3.8` if the library does not recognize a new Claude model name. * Update `tokenizers/index.ts` (or equivalent) to call `countTokens` whenever `tenant === 'claude'`. * Test in a dummy Claude chat interface (or mock DOM) to confirm the count matches server‐side counts.
