Anthropic Messages

POST http://127.0.0.1:8080/v1/messages. The SDKs’ base URL is http://127.0.0.1:8080, without /v1.

curl

curl http://127.0.0.1:8080/v1/messages -H 'content-type: application/json' -H 'anthropic-version: 2023-06-01' \
  -d '{"model": "pooled", "max_tokens": 200, "messages": [{"role": "user", "content": "Hi"}]}'

Python

import anthropic

client = anthropic.Anthropic(base_url="http://127.0.0.1:8080", api_key="pooled")
with client.messages.stream(
    model="pooled",
    max_tokens=500,
    messages=[{"role": "user", "content": "What is a KV cache?"}],
) as stream:
    for text in stream.text_stream:
        print(text, end="")

Parameters

Parameter	Behavior
max_tokens	Required, as in Anthropic’s API. Above 65536 is capped, not refused. Always capped by the context left.
system	A string or text blocks.
messages	user and assistant turns with text, tool_use, tool_result and thinking blocks.
temperature	0 to 1. 0 is greedy. Absent: the room’s own setting (0.8 if you set only top_k).
top_k	1 or more; above 64 is capped to 64. Default 40 when you set temperature.
stop_sequences	Up to 4, 64 characters each.
thinking	See Thinking.
tools, tool_choice	See Tool calling.
output_config.format, output_format	json_schema. See Structured output.
output_config.effort	Accepted.
top_p	Accepted and ignored.

Thinking

thinking	Effect
{"type": "enabled", "budget_tokens": N}	Thinking on. Once the reasoning reaches N tokens, the room closes the thinking block and the rest of max_tokens goes to the answer.
{"type": "adaptive"}	Thinking on, no budget.
{"type": "disabled"} or absent	Thinking off.
"display": "omitted"	Thinking runs, but the block comes back with empty text.

The thinking block’s signature carries the reasoning itself (a pooled1. value). A client that sends the block back gives the room its exact reasoning, and the bridge stores nothing.

Response

stop_reason is one of:

• end_turn: the model finished.
• stop_sequence: it hit one of stop_sequences.
• tool_use: the answer ends with tool calls.
• max_tokens: it reached max_tokens.
• model_context_window_exceeded: the room’s context filled up.

usage.cache_read_input_tokens counts prompt tokens the room already held. As in Anthropic’s API, usage.input_tokens leaves those out.

Streaming

Standard Messages events: message_start, content block start, delta and stop events (text_delta, thinking_delta, signature_delta, input_json_delta), message_delta and message_stop.

While the request waits in the queue or the room prefills, the bridge sends a ping event after every 10 s of silence. If something fails after the stream started, you get an error event.

Skipped and refused

Image and document blocks, in user turns or tool_results, become a short note: [image omitted: this model reads text only].

Skipped, with a logged warning: Anthropic’s own server tools (web search, bash, text editor and so on) and their result blocks.

A 400:

• search_result blocks
• a non-empty mcp_servers (connect MCP servers in your client instead)

/v1/messages/count_tokens is 404.

Errors use Anthropic’s shape, {"type": "error", "error": {"type", "message"}}, with 529 where OpenAI uses 429 and 503. See status codes.

Tip

GET /v1/models answers in Anthropic’s shape when the request has an anthropic-version header, or an x-api-key header and no Authorization header.