siliconflow/anthropic-adapter

siliconflow/anthropic-adapter

Language: Python

License: MIT

Stars: 1

Forks: 0

Open issues: 0

Created: 2026-08-19T06:22:32Z

Pushed: 2026-08-19T08:47:15Z

Default branch: main

Fork: no

Archived: no

README:

anthropic-adapter

[English](README.md) | [简体中文](README_CN.md)

A lightweight sidecar that translates the Anthropic Messages API (/v1/messages) into the OpenAI Chat Completions API (/v1/chat/completions), so any OpenAI-compatible inference backend (vLLM, sglang, TGI, etc.) can serve Anthropic-protocol clients — including tool calls, thinking/reasoning, and streaming.

Features

Anthropic protocol (`/v1/messages`)

• Text generation, streaming and non-streaming
• Tool calls: tools with input_schema, tool_choice, tool_result blocks, assistant tool_use blocks
• Thinking/reasoning: maps reasoning_content from the backend to Anthropic thinking content blocks; supports thinking.type to toggle enable_thinking via chat_template_kwargs
• System prompt, stop_sequences, temperature, top_p, top_k
• Proper Anthropic SSE event sequence with block state machine (thinking -> text -> tool_use), including thinking_delta, text_delta, and input_json_delta events
• Correct stop_reason mapping (end_turn, max_tokens, tool_use)

OpenAI passthrough (`/v1/chat/completions`, `/v1/models`)

• Streaming and non-streaming passthrough to backend, so OpenAI-protocol clients still work behind the same port

Quick Start

Docker

docker run -d \
-p 8080:8080 \
-e BACKEND_URL=http://127.0.0.1:8000 \
-e BACKEND_API_KEY=your-key \
-e DEFAULT_MODEL=your-model-name \
ghcr.io/chensl139-ok/anthropic-adapter:2.0.0

Environment Variables

| Variable | Default | Description | |---|---|---| | BACKEND_URL | http://127.0.0.1:8000 | OpenAI-compatible backend URL | | BACKEND_API_KEY | _(empty)_ | API key forwarded as Authorization: Bearer | | DEFAULT_MODEL | _(empty)_ | Fallback model name if request doesn't specify one | | PORT | 8080 | Listen port |

Examples

# Text generation (non-streaming)
curl http://localhost:8080/v1/messages \
-H "Content-Type: application/json" \
-H "x-api-key: any" \
-H "anthropic-version: 2023-06-01" \
-d '{
"model": "your-model",
"max_tokens": 128,
"messages": [{"role": "user", "content": "Hello!"}]
}'

# Streaming
curl http://localhost:8080/v1/messages \
-H "Content-Type: application/json" \
-H "x-api-key: any" \
-H "anthropic-version: 2023-06-01" \
-d '{
"model": "your-model",
"max_tokens": 128,
"stream": true,
"messages": [{"role": "user", "content": "Hello!"}]
}'

# Tool calls
curl http://localhost:8080/v1/messages \
-H "Content-Type: application/json" \
-H "x-api-key: any" \
-H "anthropic-version: 2023-06-01" \
-d '{
"model": "your-model",
"max_tokens": 256,
"tools": [{
"name": "get_weather",
"description": "Get weather for a city",
"input_schema": {
"type": "object",
"properties": {"city": {"type": "string"}},
"required": ["city"]
}
}],
"messages": [{"role": "user", "content": "What is the weather in Beijing?"}]
}'

# Disable thinking
curl http://localhost:8080/v1/messages \
-H "Content-Type: application/json" \
-H "x-api-key: any" \
-H "anthropic-version: 2023-06-01" \
-d '{
"model": "your-model",
"max_tokens": 128,
"thinking": {"type": "disabled"},
"messages": [{"role": "user", "content": "Hello!"}]
}'

Sidecar Example (GPU Function / K8s)

Add as a sidecarContainers entry in your pod spec:

sidecarContainers:
- name: anthropic-adapter
image: ghcr.io/chensl139-ok/anthropic-adapter:2.0.0
ports:
- containerPort: 8080
portName: anthropic
env:
- name: BACKEND_URL
value: 'http://127.0.0.1:8000'
- name: DEFAULT_MODEL
value: 'your-model-name'
readinessProbe:
httpGet:
path: /health
port: 8080
initialDelaySeconds: 5
periodSeconds: 10

Then point your gateway entrypointPort to 8080.

Protocol Mapping

| Anthropic | OpenAI | |---|---| | system field | role: system message | | content blocks (text) | content string | | tools[].input_schema | tools[].function.parameters | | tool_choice: auto / any / tool | auto / required / {type: function, function: {name}} | | tool_result content blocks | role: tool messages with tool_call_id | | assistant tool_use blocks | tool_calls array | | thinking.type: disabled | chat_template_kwargs: {enable_thinking: false} | | thinking.type: enabled | chat_template_kwargs: {enable_thinking: true} | | stop_sequences | stop | | top_k | top_k | | reasoning_content (response) | thinking content block | | tool_calls (response) | tool_use content blocks | | finish_reason: stop / length / tool_calls | stop_reason: end_turn / max_tokens / tool_use |

License

MIT