Anthropic integration

Overview

Anthropic builds Claude – the AI assistant family (Claude 3.5 Sonnet, Haiku, Opus). The Automize Anthropic connector covers the Messages API (chat with optional tool use + vision), message batching for offline workloads, and model listing.

• Regions: Anthropic's API is global – api.anthropic.com. For enterprise data residency, use AWS Bedrock or Google Vertex AI (separate connectors, not currently shipped).
• Plans: Pay-as-you-go per token. Higher usage tiers unlock larger rate limits.

Setting up the connection

Auth type: API Key.

1. Sign in to console.anthropic.com → API keys → Create key.
2. Copy the key (starts with sk-ant-).
3. In Automize, open Settings → Connectors → Anthropic → Add connection. Paste the key.

Find your credentials at https://console.anthropic.com/settings/keys.

Rate limits

Token-bucket per minute + per day, scaling by usage tier. Connector retries 429 with Retry-After. For high-volume offline workloads use the Batch API (50% discount, 24h SLA).

Data model

Anthropic's primitives:

    Models – claude-3-5-sonnet-latest, claude-3-5-haiku-latest, claude-3-opus-20240229
    Messages – request-response chat (with optional streaming)
    Batches – async groups of messages (cheaper, slower)
    Tool use – Claude can call functions you define + return results

Available operations

6 operations available. Click any row to jump to its detail.

Operations

Count Tokens

Count tokens in a message

Tips

• Returns token count for the message payload. Useful for cost estimation + context-window planning.

Create Message Batch

Create a batch of messages

Tips

• requests: array of message-creation payloads (each with custom_id for correlation).
• Returns batch_id. Process is async – poll Get Batch until status=ended.
• 50% cheaper than real-time. 24h SLA.

Create Message

Send a message to Claude

Tips

• model: e.g. 'claude-3-5-sonnet-latest' or 'claude-3-5-haiku-latest' (cheaper/faster).
• messages: [{role: 'user'|'assistant', content: ...}]. system is separate (top-level param, NOT in messages).
• max_tokens is REQUIRED (unlike OpenAI). Typical 1024-4096.
• tools: function schemas. Claude emits tool_use blocks; respond with tool_result blocks.
• Vision: content can be [{type:'text', text:'...'}, {type:'image', source:{type:'base64', media_type:'image/jpeg', data: '...'}}].

Errors

• invalid_request_error – Often max_tokens missing or messages array malformed (must alternate user/assistant).
• overloaded_error – Anthropic-side load spike. Retry with backoff.

Get Batch

Get batch status

Tips

• batch_id. Returns processing_status, request_counts, results_url (when ended).

List Batches

List message batches

No input parameters.

Tips

• Your batches. Filter by processing_status.

List Models

List available Claude models

No input parameters.

Tips

• Returns models available to your account + their context window sizes.

FAQ

Claude vs GPT-4 – when?

Claude has 200k context, strong reasoning, JSON output reliability. GPT-4o is fast + has native vision/voice/image gen. Use Claude for long-form analysis + reasoning; GPT-4o for multimodal.

Vision – how?

Pass image as base64 in content array: [{type:'image', source:{type:'base64', media_type:'image/jpeg', data:'...'}}, {type:'text', text:'What's in this?'}]. Or use URL source (must be publicly accessible).

System prompt – where?

Top-level `system` param, NOT in messages. Claude treats it as a persistent role. Multi-paragraph system prompts work well.

Tool use – how?

Pass tools array. Claude returns content blocks with type='tool_use'. You execute the tool, send back a user message with type='tool_result' content. Claude resumes.

Stop sequences vs max_tokens?

max_tokens is a hard cap. stop_sequences (array of strings) ends generation when matched. Use both for fine control.

Related connectors

• Openai
• Azure-openai
• Google-ai
• Aws-textract