openova/platform/anthropic-adapter

Anthropic Adapter

OpenAI-compatible proxy for Anthropic Claude API. Application Blueprint (see docs/ARCHITECTURE.md §4.6). Lets Apps written against the OpenAI SDK call Anthropic Claude with no code change. Pairs with the LLM Gateway in bp-cortex.

Status: Accepted | Updated: 2026-04-27

---

Overview

Anthropic Adapter provides an OpenAI-compatible API layer that translates requests to the Anthropic Claude API format, enabling tools like Claude Code to work with internal models.

flowchart LR
    subgraph Adapter["Anthropic Adapter"]
        Translate[Request Translator]
        Stream[Stream Handler]
    end

    ClaudeCode[Claude Code] -->|OpenAI Format| Adapter
    Adapter -->|Anthropic Format| Claude[Claude API]
    Adapter -->|OpenAI Format| Internal[Internal LLM]

    Claude --> Adapter
    Internal --> Adapter
    Adapter --> ClaudeCode

---

Why Anthropic Adapter?

Feature	Benefit
API translation	OpenAI ↔ Anthropic format
Claude Code support	Use internal models with Claude Code
Streaming	Real-time response translation
Model routing	Route to Claude or internal LLM

---

Use Cases

Use Case	Description
Claude Code + internal LLM	Use Claude Code with self-hosted models
API compatibility	Anthropic clients on OpenAI backends
Model switching	Seamless backend switching

---

Configuration

Deployment

apiVersion: apps/v1
kind: Deployment
metadata:
  name: anthropic-adapter
  namespace: ai-hub
spec:
  replicas: 2
  template:
    spec:
      containers:
        - name: adapter
          image: harbor.<location-code>.<sovereign-domain>/ai-hub/anthropic-adapter:latest
          ports:
            - containerPort: 8000
          env:
            - name: BACKEND_TYPE
              value: "openai"  # or "anthropic"
            - name: BACKEND_URL
              value: "http://vllm.ai-hub.svc:8000/v1"
            - name: BACKEND_API_KEY
              valueFrom:
                secretKeyRef:
                  name: adapter-secrets
                  key: backend-api-key
            - name: DEFAULT_MODEL
              value: "qwen3-32b"
          resources:
            requests:
              cpu: 100m
              memory: 256Mi

---

API Translation

Anthropic → OpenAI

Anthropic	OpenAI
messages[].content (list)	messages[].content (string)
max_tokens	max_tokens
system (top-level)	messages[0].role: system
stream: true	stream: true

Request Translation

# Anthropic format (input)
{
    "model": "claude-3-opus",
    "max_tokens": 4096,
    "system": "You are helpful.",
    "messages": [
        {"role": "user", "content": "Hello"}
    ]
}

# OpenAI format (translated)
{
    "model": "qwen3-32b",
    "max_tokens": 4096,
    "messages": [
        {"role": "system", "content": "You are helpful."},
        {"role": "user", "content": "Hello"}
    ]
}

---

Claude Code Configuration

# Set Claude Code to use adapter
export ANTHROPIC_API_KEY="your-adapter-key"
export ANTHROPIC_BASE_URL="http://anthropic-adapter.ai-hub.svc:8000"

# Claude Code will now use internal LLM
claude-code "Explain this code..."

---

Implementation

# proxy.py
from fastapi import FastAPI, Request
from fastapi.responses import StreamingResponse
import httpx

app = FastAPI()

@app.post("/v1/messages")
async def messages(request: Request):
    body = await request.json()

    # Translate Anthropic → OpenAI format
    openai_body = translate_to_openai(body)

    # Forward to backend
    async with httpx.AsyncClient() as client:
        if body.get("stream"):
            return StreamingResponse(
                stream_response(client, openai_body),
                media_type="text/event-stream"
            )
        else:
            response = await client.post(
                f"{BACKEND_URL}/chat/completions",
                json=openai_body
            )
            return translate_to_anthropic(response.json())


def translate_to_openai(anthropic_body: dict) -> dict:
    messages = []

    # Move system to first message
    if "system" in anthropic_body:
        messages.append({
            "role": "system",
            "content": anthropic_body["system"]
        })

    # Convert message content
    for msg in anthropic_body.get("messages", []):
        content = msg["content"]
        if isinstance(content, list):
            # Flatten content blocks
            content = " ".join(
                block.get("text", "")
                for block in content
                if block.get("type") == "text"
            )
        messages.append({"role": msg["role"], "content": content})

    return {
        "model": DEFAULT_MODEL,
        "messages": messages,
        "max_tokens": anthropic_body.get("max_tokens", 4096),
        "stream": anthropic_body.get("stream", False)
    }

---

Streaming Translation

async def stream_response(client, openai_body):
    async with client.stream(
        "POST",
        f"{BACKEND_URL}/chat/completions",
        json=openai_body
    ) as response:
        async for line in response.aiter_lines():
            if line.startswith("data: "):
                data = json.loads(line[6:])
                # Translate to Anthropic SSE format
                anthropic_event = translate_sse(data)
                yield f"event: content_block_delta\ndata: {json.dumps(anthropic_event)}\n\n"
        yield "event: message_stop\ndata: {}\n\n"

---

Monitoring

Metric	Query
Request count	adapter_requests_total
Latency	adapter_request_duration_seconds
Backend errors	adapter_backend_errors_total
Stream duration	adapter_stream_duration_seconds

---

Consequences

Positive:

• Claude Code with internal models
• API format translation
• Streaming support
• Easy model switching

Negative:

• Feature parity limitations
• Translation overhead
• Some Anthropic features unsupported

---

Part of OpenOva