Guides

Anthropic Messages SDK

Drop-in for the Anthropic Python/TS packages: same Messages shape, Rodium key, Claude models.

POSThttps://api.rodiumai.io/v1/messages

Prefer this path when your app already uses the Anthropic SDK. OpenAI-compatible chat still works for Claude via /v1/chat/completions.

Auth & base URL

Pass api_key=rd_sk_… (SDK sends x-api-key). Bearer Authorization is also accepted. Base URL: https://api.rodiumai.io

Anthropic SDK vs OpenAI SDK

OpenAI clients use base_url https://api.rodiumai.io/v1 and chat.completions. Anthropic clients use base_url https://api.rodiumai.io (no /v1) and messages.create. Same rd_sk_… secret; never send Anthropic’s server key to clients.

Install

…

First message

…

Streaming (native Anthropic SSE)

…

Claude-first MVP

Only anthropic/* models are routed on /v1/messages today. Multi-provider Messages bridging is planned later.

Request parameters

Parameter	Type	Required	Description
model	string	Required	Claude catalogue id (e.g. anthropic/claude-sonnet-4-6). Only anthropic/* models are served on this endpoint. POST /v1/messages/count_tokens takes the same body and returns input_tokens for free.
messages	array	Required	Anthropic Messages array (role + content blocks or text).
max_tokens	integer	Required	Maximum output tokens (required by Anthropic Messages).
system	string \| array	Optional	Optional system prompt.
stream	boolean	Optional	When true, returns native Anthropic SSE (message_start, content_block_delta, …).
temperature	number	Optional	Sampling temperature when supported.
tools	array	Optional	Anthropic tool definitions; tool_use / tool_result blocks round-trip unchanged.

API reference: POST /v1/messages

Scope of /v1/messages

Only Claude models (anthropic/*) are served here; RodiumAI may route them through Anthropic, Amazon Bedrock or Azure without changing the response shape. Other models are not available on this endpoint: call them with /v1/chat/completions.
On routes without a native Messages API (Amazon Bedrock), the gateway translates the request: client tools, tool_use / tool_result turns and images are kept. Provider-hosted tools (web search, code execution) only run on native Anthropic routes; elsewhere they are skipped.
Smart aliases (rodiumai/smart, rodium/*) and custom models are chat-completions only.
Authenticate with x-api-key (what the Anthropic SDK sends) or Authorization: Bearer. anthropic-version is optional.
Set max_tokens: it is required by the Messages format and bounds the cost. speed values other than standard are not forwarded, so requests run at standard speed and price.
Searches made by Anthropic's hosted web search tool (usage.server_tool_use.web_search_requests) are billed per call on top of tokens, see Pricing.

POST /v1/messages/count_tokens

Anthropic-compatible token counting, so clients such as Claude Code can size their context window. Send the same body as /v1/messages (model is required; max_tokens is not); the answer is {"input_tokens": N}. Same authentication as /v1/messages (x-api-key or Authorization: Bearer). The call is free: nothing is reserved or billed.

…

input_tokens is the gateway's own estimate (the one used to size pre-flight holds), not the provider's tokenizer: use it to size prompts, not to predict the bill exactly. Errors use the Anthropic format; an invalid key returns 401 authentication_error.

Errors

Body validation and streaming errors use Anthropic's {"type": "error", "error": {…}} format; errors raised before the call (authentication, balance, quota, rate limit) use the OpenAI-style envelope. Details: API errors.