Skip to main content

Anthropic Adapter Implementation

⬅️ Back to Features Catalog

What It Does​

The Anthropic Adapter Implementation is a highly specialized protocol translator built natively into the proxy. It allows any OpenAI-compatible client to seamlessly interface with Anthropic's Claude API. It fundamentally bridges the gap between OpenAI's monolithic messages schema and Anthropic's distinct prompt engineering requirements.

How It Works​

Anthropic's Messages API enforces several strict contracts that the OpenAI API does not. The adapter normalizes these differences at the network edge:

  1. System Prompt Extraction: OpenAI allows multiple system messages anywhere in the array. Anthropic requires a top-level system string. The adapter extracts all system roles, concatenates them safely, and lifts them to the root JSON object.
  2. Strict Alternation (User/Assistant): Claude rejects requests where two user messages occur consecutively. The adapter detects this and merges consecutive contents, separated by newlines, into a single valid message block.
  3. Multi-Content Block Normalization: Anthropic streams back complex events like message_start, content_block_start, content_block_delta, and message_delta. The adapter's asynchronous generator converts these proprietary events back into standard OpenAI choices[0].delta.content Server-Sent Events (SSE) chunks on the fly.

View diagram on GitHub mobile 📱 -->

Performance Profile​

  • Execution Speed: Schema mapping adds nearly zero latency, operating in O(N) relative to the number of message objects in the array.
  • Overhead: Uses zero-allocation techniques to prevent massive payload duplication in memory.

Configuration Flags​

The adapter engages automatically when the proxy detects an Anthropic target URL.

Environment VariableDescriptionLinked Deployment Guide
ANTHROPIC_VERSIONSets the anthropic-version API header. Default: 2023-06-01.View in deployment.md

Critical Logic & Edge Cases​

  • Tool Calling (Function Calling) Compatibility: The adapter is fully aware of Anthropic's tool-use XML and JSON specifications, successfully mapping OpenAI's tools array into Anthropic's format, and parsing Claude's tool execution requests back into standard OpenAI JSON-RPC payloads.

FAQ​

Q: Can I use Claude 3.5 Sonnet directly from my existing OpenAI SDK? A: Yes! Set the model string in your SDK to claude-3-5-sonnet-20240620, point your base URL to the proxy, and the proxy will automatically route and translate the request to Anthropic.

Q: Does Anthropic's SSE stream break the sliding-window buffer? A: No. The adapter normalizes Anthropic's stream into standard chunks before it passes them into the SSE Rehydration Buffer, ensuring that PII de-masking works flawlessly across providers.

Plainspeak​

This feature specifically handles the unique, strict conversational rules required by Anthropic's Claude AI.

Anthropic is extremely picky about how a conversation is formatted (for example, it requires exactly alternating "User" and "Assistant" messages). If your application sends messages out of order, Anthropic will reject them. This adapter acts as a smart editor, automatically reformatting and fixing your message history in real-time so that Anthropic accepts it without complaints.

See the following test file for reference implementations and edge-case testing: tests/test_provider_adapters.py.