Anthropic Messages

---
title: Anthropic Messages
description: Use Anthropic Messages through Relyven with system prompts, tools, thinking, caching, and streaming.
lastUpdated: 2026-07-29
---

## Endpoint and authentication

```text
POST https://api.relyven.com/v1/messages
```

```bash
curl https://api.relyven.com/v1/messages \
  -H "x-api-key: $RELYVEN_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "YOUR_CLAUDE_MODEL_ID",
    "max_tokens": 256,
    "messages": [{"role":"user","content":"Reply with OK."}]
  }'
```

`max_tokens` is required by the Messages API. The selected model must be within the key's allowed scope.

## System instructions

Messages uses a top-level `system` field. Do not add a `role: "system"` entry to the `messages` array:

```json
{
  "model": "YOUR_CLAUDE_MODEL_ID",
  "max_tokens": 256,
  "system": "Answer concisely.",
  "messages": [{"role":"user","content":"Explain health checks."}]
}
```

Relyven does not intentionally inject or override the system prompt.

## Streaming

Set `"stream": true` to receive SSE. Handle Anthropic event types such as `message_start`, `content_block_delta`, `message_delta`, and `message_stop`; do not parse them as OpenAI streaming chunks.

## Tool use

Declare tools and their input schemas in `tools`. When Claude returns a `tool_use` content block, execute the tool and send a matching `tool_result` in a later user message.

```json
{
  "name": "get_weather",
  "description": "Get weather for a city",
  "input_schema": {
    "type": "object",
    "properties": {"city": {"type": "string"}},
    "required": ["city"]
  }
}
```

Always validate model-generated tool arguments. Add confirmation and idempotency controls for operations with side effects.

## Prompt caching

Models with prompt-caching support can accept `cache_control` on system, message content, tool definitions, and other supported blocks. Availability, minimum cacheable content, and billing depend on the model and model provider.

Relyven does not intentionally remove `cache_control`; this does not imply identical caching support across every Claude model.

## Extended thinking

Models that provide the capability can accept a `thinking` configuration. Its constraints interact with `max_tokens`, tool use, and model versions; follow Anthropic's requirements for the selected model.

Relyven does not intentionally change `thinking` or reduce its configured level. Unsupported combinations may still return a model-provider `400`.

## Troubleshooting

- `400`: check `max_tokens`, top-level `system`, content blocks, and model capabilities.
- `401/403`: check `x-api-key`, model permission, and the base URL.
- Stream parsing failure: use Anthropic SSE event types, not OpenAI chunk shapes.
- Broken tool loop: verify that `tool_use.id` matches `tool_result.tool_use_id`.

Official references: [Messages API](https://docs.anthropic.com/en/api/messages) · [Streaming](https://docs.anthropic.com/en/docs/build-with-claude/streaming) · [Tool use](https://docs.anthropic.com/en/docs/build-with-claude/tool-use/overview) · [Prompt caching](https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching) · [Extended thinking](https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking)

Endpoint and authentication

POST https://api.relyven.com/v1/messages

curl https://api.relyven.com/v1/messages \
  -H "x-api-key: $RELYVEN_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "YOUR_CLAUDE_MODEL_ID",
    "max_tokens": 256,
    "messages": [{"role":"user","content":"Reply with OK."}]
  }'

max_tokens is required by the Messages API. The selected model must be within the key’s allowed scope.

System instructions

Messages uses a top-level system field. Do not add a role: "system" entry to the messages array:

{
  "model": "YOUR_CLAUDE_MODEL_ID",
  "max_tokens": 256,
  "system": "Answer concisely.",
  "messages": [{"role":"user","content":"Explain health checks."}]
}

Relyven does not intentionally inject or override the system prompt.

Streaming

Set "stream": true to receive SSE. Handle Anthropic event types such as message_start, content_block_delta, message_delta, and message_stop; do not parse them as OpenAI streaming chunks.

Tool use

Declare tools and their input schemas in tools. When Claude returns a tool_use content block, execute the tool and send a matching tool_result in a later user message.

{
  "name": "get_weather",
  "description": "Get weather for a city",
  "input_schema": {
    "type": "object",
    "properties": {"city": {"type": "string"}},
    "required": ["city"]
  }
}

Always validate model-generated tool arguments. Add confirmation and idempotency controls for operations with side effects.

Prompt caching

Models with prompt-caching support can accept cache_control on system, message content, tool definitions, and other supported blocks. Availability, minimum cacheable content, and billing depend on the model and model provider.

Relyven does not intentionally remove cache_control; this does not imply identical caching support across every Claude model.

Extended thinking

Models that provide the capability can accept a thinking configuration. Its constraints interact with max_tokens, tool use, and model versions; follow Anthropic’s requirements for the selected model.

Relyven does not intentionally change thinking or reduce its configured level. Unsupported combinations may still return a model-provider 400.

Troubleshooting

400: check max_tokens, top-level system, content blocks, and model capabilities.
401/403: check x-api-key, model permission, and the base URL.
Stream parsing failure: use Anthropic SSE event types, not OpenAI chunk shapes.
Broken tool loop: verify that tool_use.id matches tool_result.tool_use_id.

Official references: Messages API · Streaming · Tool use · Prompt caching · Extended thinking