@gotong/llm-anthropic

Anthropic Claude LlmProvider for Gotong. Plug into @gotong/llm's LlmAgent.

Install

pnpm add @gotong/llm-anthropic @anthropic-ai/sdk
# @anthropic-ai/sdk is a peer dependency — install the version you want

Use

import { LlmAgent } from '@gotong/llm'
import { AnthropicProvider } from '@gotong/llm-anthropic'

const provider = new AnthropicProvider({
  // apiKey: '...'         // defaults to process.env.ANTHROPIC_API_KEY
  // defaultModel: '...'   // defaults to 'claude-opus-4-7'
  // defaultMaxTokens: ... // defaults to 1024
})

const writer = new LlmAgent({
  id: 'writer',
  capabilities: ['draft'],
  provider,
  system: 'You write one short sentence and stop.',
})

Translation rules

• Request → messages.create({ model, max_tokens, messages, system?, temperature? })
• Response text = concat of all content[] blocks with type === 'text'
• stop_reason: 'end_turn' | 'stop_sequence' → 'end_turn'; 'max_tokens' → 'max_tokens'; anything else → 'error'
• usage.input_tokens / output_tokens → inputTokens / outputTokens

Note on Claude Opus 4.x (thinking models)

Opus 4.x ("thinking" model family) does not accept temperature / top_p / top_k; the API returns 400 if any of them is set. The provider detects these models by prefix (claude-opus-4-*) and silently drops any user-supplied temperature before sending — adaptive thinking takes over. If you need a specific temperature, target a non-thinking model (e.g. claude-sonnet-4-6, claude-haiku-4-5).

License

MIT