@sisu-ai/adapter-anthropic

v10.0.0

Published

2 months ago

Connect Sisu to Anthropic models with first-class tool calling, streaming, and vision input support.

0High
0Medium
0Low

lejahmie

sisu ai ai-agent agentic adapter anthropic

@sisu-ai/adapter-anthropic

Connect Sisu to Anthropic models with first-class tool calling, streaming, and vision input support.

Setup

npm i @sisu-ai/adapter-anthropic

Env: API_KEY
Optional base URL: BASE_URL
Optional debug: DEBUG_LLM=1 to log redacted request/response summaries on errors

Transport and compatibility

The adapter now uses the official @anthropic-ai/sdk Messages API transport.
Public Sisu behavior is intentionally preserved: message normalization, tool-call mapping, streaming token/assistant_message events, and usage mapping.
Existing app code should continue to work unchanged. If you relied on low-level fetch stubs in tests, prefer Response-compatible mocks or SDK-boundary mocks.

Usage

import { anthropicAdapter, anthropicEmbeddings } from '@sisu-ai/adapter-anthropic';

const model = anthropicAdapter({ model: 'claude-3-5-sonnet-20240620' });
// with a self-hosted proxy
const model = anthropicAdapter({ model: 'claude-3-opus-20240229', baseUrl: 'https://api.anthropic.com' });

const embeddings = anthropicEmbeddings({
  baseUrl: process.env.EMBEDDINGS_BASE_URL!,
  model: process.env.EMBEDDINGS_MODEL!,
  apiKey: process.env.EMBEDDINGS_API_KEY,
});

Embeddings

Anthropic does not provide a native embeddings API.
anthropicEmbeddings(...) is a convenience helper for Anthropic-centered apps that use a compatible third-party embeddings endpoint.
You must pass baseUrl and model explicitly; there are no Anthropic-specific defaults.

Tools

Sends tools in Anthropic format (name, description, input_schema).
Sends tool_choice only when tools exist (Anthropic rejects tool_choice without tools).
Maps assistant tool calls to { id, name, arguments } for the middleware loop.
Maps tool results into Anthropic tool_result blocks and back.

Vision

Supports mixed text + image user messages.
Accepts content parts (content / contentParts) with { type: 'text', text } and { type: 'image_url', image_url }.
Accepts convenience fields: image, image_url, images, image_urls.
Image inputs may be:
- data:image/...;base64,... URLs
- raw base64 payloads
- remote http(s) URLs (fetched and converted to base64 before API call)

Streaming

Supports server‑sent events from the Messages API.
Emits { type: 'token', token } events for content deltas and a final { type: 'assistant_message', message }.

Options

anthropicAdapter({
  model: 'claude-3-5-sonnet-20240620',
  apiKey?: string,            // default: API_KEY
  baseUrl?: string,           // default: https://api.anthropic.com (or BASE_URL)
  anthropicVersion?: string,  // default: 2023-06-01
  timeout?: number,           // default: 60000 ms
  maxRetries?: number,        // default: 3 (with backoff; 4xx except 429 are not retried)
});

Message mapping

System messages → system string (joined if multiple system messages appear)
User messages → text blocks and (optionally) image blocks with Anthropic base64 image sources
Assistant messages with tool calls → tool_use blocks with { id, name, input }
Tool messages → user tool_result blocks with { tool_use_id | name, content }

Usage reporting

The adapter maps usage.input_tokens and usage.output_tokens to { promptTokens, completionTokens, totalTokens }.

Debugging

DEBUG_LLM=1 prints sanitized payloads and error bodies.
Combine with your app logger’s LOG_LEVEL=debug to see middleware events.

Community & Support

Discover what you can do through examples or documentation. Check it out at https://github.com/finger-gun/sisu. Example projects live under examples/ in the repo.