Anthropic Messages API¶

POST /v1/messages serves the Anthropic Messages API on the same listener as the OpenAI and Ollama APIs. Reasoning comes back as thinking blocks, text as text blocks and parsed tool calls as tool_use blocks.

curl http://127.0.0.1:8453/v1/messages -H 'content-type: application/json' -d '{
  "model": "unsloth/Qwen3.8-27B-GGUF",
  "max_tokens": 256,
  "system": "Be terse.",
  "messages": [{"role": "user", "content": "Why is the sky blue?"}]
}'

With the Anthropic Python SDK:

from anthropic import Anthropic

client = Anthropic(base_url="http://127.0.0.1:8453", api_key="unused")
msg = client.messages.create(
    model="unsloth/Qwen3.8-27B-GGUF",
    max_tokens=256,
    messages=[{"role": "user", "content": "Why is the sky blue?"}],
)
print(msg.content)

No anthropic-version header is required. A key may be sent as x-api-key or Authorization: Bearer.

Request¶

field	default	notes
`model`	required	a `/v1/models` id
`max_tokens`	required
`messages`	required	`content` a string or an array of blocks
`system`	none	a string or `text` blocks (joined with newlines)
`tools`	`[]`	`{name, description?, input_schema, cache_control?}`
`stream`	`false`
`temperature`, `top_p`, `top_k`	model default	resolved as on the OpenAI routes; see Sampling

Ignored: stop_sequences, tool_choice, thinking, metadata, service_tier. There is no seed; a sampled request draws a fresh one.

Content blocks¶

block	mapping
`text`	text
`image` with `source.type: "base64"`	an image, on a model that accepts images (else 400 `unsupported_input`); URL sources are refused
`tool_use`	a prior call in assistant history
`tool_result`	a tool result; `content` a string or text blocks. `is_error` is not read
`thinking`	prior reasoning in assistant history

Other block types are skipped.

Prompt caching (`cache_control`)¶

A cache_control breakpoint on a tool, a system block or a content block pins the prompt prefix up to that turn in the KV cache for its TTL, so pressure eviction cannot drop it.

{"system": [{"type": "text", "text": "<long shared preamble>",
             "cache_control": {"type": "ephemeral", "ttl": "1h"}}]}

type must be ephemeral; ttl is 5m (default) or 1h.
At most 4 breakpoints per request; a 1h breakpoint may not follow a 5m one.
Not allowed on thinking blocks or empty text blocks.
Pins together hold at most --pin-budget-pct of the KV pool (default 50) and yield, oldest first, under pressure.
Prompts carrying images pin nothing.

Response¶

{
  "id": "msg_superfluid_12", "type": "message", "role": "assistant", "model": "Qwen3-4B",
  "content": [
    {"type": "thinking", "thinking": "...", "signature": ""},
    {"type": "text", "text": "..."},
    {"type": "tool_use", "id": "toolu_418", "name": "get_weather", "input": {"city": "Oslo"}}
  ],
  "stop_reason": "tool_use", "stop_sequence": null,
  "usage": {"input_tokens": 24, "cache_creation_input_tokens": 96,
            "cache_read_input_tokens": 0, "output_tokens": 58}
}

stop_reason is end_turn, max_tokens or tool_use. A reply cut by max_tokens may hold a truncated tool_use; do not run it.
cache_read_input_tokens is the prefix served from the prefix cache; cache_creation_input_tokens is what this request's breakpoints pinned beyond that.
A tool block the parser could not read is returned as a text block with the raw text.
signature is always empty.

Streaming¶

stream: true emits the standard event sequence: message_start, then content_block_start / content_block_delta / content_block_stop per block, message_delta with stop_reason and usage, and message_stop.

Text and thinking deltas are sent once per scheduler tick.
A tool call is emitted when it closes, as one input_json_delta carrying the whole arguments.
Keep-alives are SSE comment lines; there are no ping events.
An error after the stream began is event: error.

Tools and thinking¶

Tools are rendered through the model's own chat template, as OpenAI function tools are. Decoding is not grammar-constrained on this route (there is no tool_choice). Return a tool_use block in the assistant turn and answer it with a tool_result block in the next user message. See Tool calling.

Whether the model reasons is its template's default; this route passes no template variables. Use the OpenAI route's chat_template_kwargs to switch thinking off.

Errors¶

{"type": "error", "error": {"type": "invalid_request_error", "message": "..."}}

Every daemon-side refusal on this route is a 400 (invalid_request_error); a worker failure is a 500 (api_error). Refusals made before the handler runs (401, 403, 404, 413, 429) use the OpenAI error envelope.

Not supported¶

stop_sequences, tool_choice, thinking configuration, metadata, service_tier
image URLs, audio blocks, documents, redacted_thinking
token-by-token input_json_delta, ping events
message batches, token counting, and the Anthropic Files API