Anthropic Compatible API

One API Pro uses the relay/adaptor/anthropic package to translate OpenAI-style Chat Completions requests (/v1/chat/completions) into Anthropic Messages, and to consume Anthropic SSE streaming events. This page describes the Anthropic request / response and streaming-event conventions so clients can talk to Claude-style models through the gateway.

The native /v1/messages endpoint is currently served by /v1/chat/completions: when a claude-* model is requested, the body is sent upstream in Anthropic Messages form, and the response is normalized into OpenAI Chat Completions shape. The Anthropic wire format documented below is the form the adapter speaks to the upstream.

Endpoint index

Endpoint	Method	Auth	Description
/v1/chat/completions (model=claude-*)	POST	Bearer Token	OpenAI Chat Completions entry, hits the anthropic adapter
/v1/models	GET	Bearer Token	Model list (OpenAI-compatible)

1. Request Format (Anthropic Messages)

Internally the adapter converts OpenAI Chat Completions requests into the Anthropic Messages shape below (relay/adaptor/anthropic/main.go::ConvertRequest).

Request body:

{
  "model": "claude-3-5-sonnet",
  "messages": [
    { "role": "user", "content": "Hello" }
  ],
  "system": "You are a helpful assistant.",
  "max_tokens": 4096,
  "temperature": 0.7,
  "top_p": 0.9,
  "top_k": 40,
  "stop_sequences": ["\n\nHuman:"],
  "stream": false,
  "tools": [
    {
      "name": "get_weather",
      "description": "Get current weather",
      "input_schema": {
        "type": "object",
        "properties": { "city": { "type": "string" } },
        "required": ["city"]
      }
    }
  ],
  "tool_choice": { "type": "auto" }
}

Fields:

Field	Type	Required	Description
model	string	yes	Model name. Common: claude-3-5-sonnet, claude-3-opus, etc.
messages	array	yes	Conversation history, see Message below
system	string	no	System prompt. A role=system entry in messages is folded into this field.
max_tokens	int	no	Max output tokens. Filled with 4096 when omitted.
temperature	float	no	Sampling temperature
top_p	float	no	Nucleus sampling
top_k	int	no	Top-K sampling
stop_sequences	array<string>	no	Custom stop sequences
stream	bool	no	Whether to stream the response
tools	array	no	Tool list, each a Tool
tool_choice	object/string	no	Tool choice; OpenAI function shape or string auto / any

Message:

{ "role": "user", "content": [{ "type": "text", "text": "Hello" }] }

Field	Type	Description
role	string	user / assistant. OpenAI's tool role is mapped to user + tool_result.
content	array	Content blocks (Content)

Content:

Field	Type	Description
type	string	text / image / tool_use / tool_result
text	string	Text payload (type=text or type=tool_result)
source	object	Image source (type=image); {type:"base64", media_type:"image/png", data:"..."}
id	string	tool_use id
name	string	tool_use tool name
input	object	tool_use arguments
tool_use_id	string	Matching tool_use.id for a tool_result
content	string	Text payload of a tool_result

Tool:

Field	Type	Description
name	string	Tool name
description	string	Tool description
input_schema	object	JSON Schema for the tool arguments; contains type / properties / required

2. Response Format (non-streaming)

{
  "id": "msg_01XYZ",
  "type": "message",
  "role": "assistant",
  "content": [
    { "type": "text", "text": "Hello! How can I help you today?" }
  ],
  "model": "claude-3-5-sonnet",
  "stop_reason": "end_turn",
  "stop_sequence": null,
  "usage": {
    "input_tokens": 12,
    "output_tokens": 18,
    "cache_read_input_tokens": 0,
    "cache_creation_input_tokens": 0
  },
  "error": { "type": "", "message": "" }
}

Fields:

Field	Type	Description
id	string	Message ID
type	string	Always message
role	string	Always assistant
content	array	Content blocks (same shape as above)
model	string	Actual model used
stop_reason	string	end_turn / stop_sequence / max_tokens / tool_use
stop_sequence	string/null	Stop sequence hit, if any
usage.input_tokens	int	Input tokens
usage.output_tokens	int	Output tokens
usage.cache_read_input_tokens	int	Cache read tokens
usage.cache_creation_input_tokens	int	Cache write tokens
error	object	Error object; empty on success

The adapter maps stop_reason into OpenAI finish_reason: end_turn / stop_sequence → stop, max_tokens → length, tool_use → tool_calls. The response is returned on /v1/chat/completions in OpenAI chat.completion shape.

3. Streaming Response (SSE)

SSE event order follows the Anthropic Messages streaming spec; clients dispatch by the event prefix:

event: message_start
data: {"type":"message_start","message":{...full message...}}

event: content_block_start
data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Hello"}}

event: content_block_stop
data: {"type":"content_block_stop","index":0}

event: message_delta
data: {"type":"message_delta","delta":{"stop_reason":"end_turn"},"usage":{"output_tokens":18}}

event: message_stop
data: {"type":"message_stop"}

Event	Key fields	Description
message_start	message.{id,model,role,usage}	Message metadata; usage is cumulative.
content_block_start	content_block.{type,text/id/name/input}	Opens a content block (text or tool_use).
content_block_delta	delta.{type,text/partial_json}	Incremental content; input_json_delta carries partial JSON for tool args.
content_block_stop	—	Closes the current content block.
message_delta	delta.{stop_reason,stop_sequence}, usage	Cumulative usage (max-merged to avoid double counting).
message_stop	—	Closes the whole message.

At /v1/chat/completions, One API Pro normalizes these events into OpenAI chat.completion.chunk shape so OpenAI-SDK clients can consume them directly.

4. Auth and Billing

• Auth: Bearer Token (i.e. sk-xxxxxxxx issued by /api/token/), identical to the OpenAI-compatible endpoints.
• Billing: Computed from OpenAI-side prompt_tokens + completion_tokens; cache_read_input_tokens / cache_creation_input_tokens are folded into prompt_tokens by model.ClaudeUsage2OpenAI.
• Retry / Fallback: Same as the OpenAI-compatible endpoints (config.RetryTimes, ErrorNext), driven by the generic relay/handler retry chain.

5. Error Response (from upstream)

{
  "type": "error",
  "error": {
    "type": "invalid_request_error",
    "message": "messages: must be non-empty"
  }
}

After relaying through /v1/chat/completions, the final body uses the standard OpenAI error shape:

{
  "error": {
    "message": "<msg> (request id: <uuid>)",
    "type": "one_api_error",
    "param": "",
    "code": "<upstream error type>"
  }
}