@weldable/integration-anthropic

Anthropic Claude API actions for Weldable.

Part of the Weldable integration library — see @weldable/integration-core for the full catalog.

Install

npm install @weldable/integration-anthropic @weldable/integration-core

@weldable/integration-core is a peer dependency and must be installed alongside this package.

Usage

import integration from '@weldable/integration-anthropic'

const action = integration.actions.find(a => a.id === 'anthropic.llm')!

const result = await action.execute(
  {
    prompt: 'Summarize the following in one sentence: Node.js is a JavaScript runtime built on V8.',
    model: 'claude-sonnet-4-6',
    system: 'You are a concise technical writer.',
  },
  ctx, // ActionContext from your Weldable-compatible host
)

console.log(result.content) // "Node.js is a V8-powered JavaScript runtime for server-side execution."

Prompt caching

Pass prompt (or system) as an array of blocks to cache stable prefix content across repeated calls within the same run:

- uses: anthropic.llm
  with:
    system: "You are a career advisor..."
    prompt:
      - text: "{{ steps.read_resume.output.markdown }}"
        cache: true
      - text: "Evaluate this job: {{ job.title }} at {{ job.company }}"

Mark stable blocks with cache: true; the last block (variable per-call content) should not be marked. On the first call within a 5-minute window, the marked blocks are written to cache (1.25× the normal input price). Subsequent calls within that window that share the same prefix read from cache at 0.1× the normal price — a 10× discount.

Check usage.cacheCreationInputTokens (> 0 on write) and usage.cacheReadInputTokens (> 0 on hit) in the step output to verify caching is active. Note: Sonnet 4.6 requires at least 1,024 tokens in the cached prefix; below that threshold the cache: true marker is silently ignored.

Contributing and releasing

See CONTRIBUTING.md in @weldable/integration-core for the development workflow and release process.

License

MIT