The Anthropic model driver for Namzu.

Install · Usage · Documentation

Implements the kernel's LLMProvider interface over the official @anthropic-ai/sdk, so nothing above the driver layer knows which vendor is answering. Installed only if you use it — the kernel has no preferred vendor and no driver is a dependency of it.

Install

pnpm add @namzu/sdk @namzu/anthropic

@namzu/sdk is a peer dependency. Install both.

Usage

import { ProviderRegistry, collectChatCompletion } from '@namzu/sdk'
import { registerAnthropic } from '@namzu/anthropic'

// Once, at startup. Module augmentation extends the config union, so the
// call below is fully type-narrowed on `type: 'anthropic'`.
registerAnthropic()

const { provider } = ProviderRegistry.create({
  type: 'anthropic',
  apiKey: process.env.ANTHROPIC_API_KEY,
  model: 'claude-sonnet-4-5-20250929',
})

const response = await collectChatCompletion(
  provider.chatStream({
    model: 'claude-sonnet-4-5-20250929',
    messages: [{ role: 'user', content: 'Hello' }],
  }),
)

chatStream is the only model entry point; a non-streaming call is that stream collected. In practice the kernel's run loop calls it and hands you events.

Signed thinking and encrypted redacted blocks are persisted with versioned adapter state plus their exact provider/model/fallback route. The same configured route replays them after restart or /resume; another model, provider, or chain member receives portable assistant/tool history without foreign native thinking metadata.

Set exactly one of apiKey or authToken. The kernel's credential vault can hold the key instead, so it never reaches the driver's config as a plain string.

Documentation

• The Anthropic driver — every configuration field, extended thinking, strict tool inputs
• Namzu docs

License

FSL-1.1-MIT, converting to MIT two years after each release.