# Anthropic Messages compatibility

Connect Messages-protocol clients to KosCompute-hosted models.

`POST /v1/messages` provides Anthropic Messages **protocol compatibility**. KosCompute-hosted models serve the request; the request is not sent to Anthropic.

Use the host-level base URL expected by Anthropic clients:

```text
https://api.koscompute.com
```

```bash
curl https://api.koscompute.com/v1/messages \
  -H "x-api-key: $KOSCOMPUTE_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen/qwen3.8-27b",
    "max_tokens": 180,
    "system": "Answer precisely.",
    "messages": [{"role": "user", "content": "What is backpressure?"}]
  }'
```

Bearer authentication is also accepted. If Bearer and `x-api-key` are both supplied, they must match.

## Messages and tools

`system` accepts text or text blocks. User and assistant content can use compatible content blocks. Function declarations are converted to local tools; `tool_use` responses and `tool_result` inputs preserve their IDs. `tool_choice` supports automatic, any/required, and named-tool forms where the selected model can satisfy them.

## Thinking

Compatible `thinking` controls are mapped to the selected model's reasoning support. A reasoning budget remains inside `max_tokens`. Unsupported or adaptive provider-specific thinking fields can be consumed for client compatibility but do not imply an identical Anthropic model implementation.

## Streaming

Streaming uses Anthropic-style event names including `message_start`, content-block start/delta/stop, `message_delta`, and `message_stop`. Tool input arrives as JSON deltas.

## Error envelope

Errors use the Anthropic envelope: `{ "type":"error", "error": { "type":..., "message":... } }`. Normal rate and concurrency limits remain `429`. Upstream transport or provider-unavailable conditions that are `502`, `503`, or `504` in OpenAI-compatible endpoints intentionally map to Anthropic-compatible `529 overloaded_error` here.