Skip to main content
MonoRelay accepts native Anthropic API format and exposes it at /v1/messages — you can point the official Anthropic SDK directly at MonoRelay without changing your request format. All request fields, streaming behavior, and beta headers are forwarded transparently to the configured Anthropic-compatible provider.

​
Endpoints

The Anthropic-compatible endpoints share the same /v1/ base path as OpenAI-compatible ones:
When configuring the Anthropic SDK with base_url, use http://localhost:8787 (without /v1/). The SDK appends /v1/messages automatically.

​
Authentication

Pass your MonoRelay access token in either of these headers: The token is verified against the MonoRelay server’s configured access_key or a valid user JWT issued after login.

​
POST /v1/messages

Send a chat message and receive an Anthropic-format response. Supports both single-turn and multi-turn conversations, streaming, vision, tool use, and extended thinking (beta).

​
GET /v1/anthropic/models

List models available through the configured Anthropic-compatible provider. Returns a model list in Anthropic response format.

​
POST /v1/messages/beta

Identical to the standard messages endpoint but also accepts the anthropic-beta header for features such as extended thinking, prompt caching, and computer use. Set the header in your request and MonoRelay forwards it upstream unchanged.

​
Request body

​
string
required
The model identifier to use, e.g. claude-opus-4-5 or a MonoRelay alias. MonoRelay resolves aliases and provider routing before forwarding.
​
integer
required
Maximum number of tokens to generate. Required by the Anthropic API; has no default.
​
array
required
Ordered list of conversation turns. Each element must have:
  • role — "user" or "assistant"
  • content — a string or an array of content blocks (text, image, tool_result, etc.)
The array must alternate roles and must begin with a user turn.
​
string or array
System prompt text. Accepts either a plain string or an array of content blocks for structured system prompts (useful with prompt caching).
​
number
Sampling temperature between 0.0 and 1.0. Controls output randomness. Omit to use the model’s default.
​
number
Nucleus sampling threshold. Only tokens in the top top_p probability mass are considered. Usually set either temperature or top_p, not both.
​
integer
Only sample from the top top_k tokens at each step. Not available on all models.
​
array
List of strings that stop generation when encountered. The model will not include the stop sequence in the response.
​
boolean
Set to true to receive a Server-Sent Events stream instead of a single JSON response. Defaults to false.
​
array
List of tool definitions the model may call. Each tool requires name, description, and a JSON Schema input_schema.
​
object
Controls tool selection behavior. Use {"type": "auto"}, {"type": "any"}, or {"type": "tool", "name": "my_tool"} to force a specific tool.

​
Response fields

​
string
Unique message identifier, e.g. msg_01XFDUDYJgAACTJgxXiHvVEF.
​
string
Always "message" for non-streaming responses.
​
string
Always "assistant".
​
array
Array of content blocks. A standard text response contains one block of {"type": "text", "text": "..."}. Tool use responses include {"type": "tool_use", ...} blocks. Extended thinking responses include {"type": "thinking", "thinking": "..."} blocks.
​
string
The model that generated the response as resolved by MonoRelay.
​
string
Why generation stopped: "end_turn", "max_tokens", "stop_sequence", or "tool_use".
​
object
Token consumption report:
  • input_tokens — prompt tokens consumed
  • output_tokens — completion tokens generated
  • cache_read_input_tokens — tokens served from prompt cache
  • cache_creation_input_tokens — tokens written to prompt cache

​
Python SDK example

Point the official Anthropic Python SDK at MonoRelay by setting base_url. No other code changes are needed.

​
curl example

​
Streaming example

Streaming responses arrive as Server-Sent Events. The sequence of event types is: message_start → content_block_start → content_block_delta (repeated) → content_block_stop → message_delta → message_stop.