Use the Anthropic SDK

Proxium speaks the Anthropic Messages API at POST /v1/messages. A Messages call can go to any vendor of your project, not only Anthropic. Proxium translates the call and the answer.

Setting	Value
Base URL	https://proxium.tech, without /v1. The SDK adds /v1/messages
API key	A virtual key of your project, from Keys in the console
Model	A model id or a tier name of your project. Route requests to models explains both

Connect the client

To connect, set the base URL and give your virtual key as the API key.

Python

client.py

import os
from anthropic import Anthropic

client = Anthropic(
    base_url="https://proxium.tech",
    api_key=os.environ["PROXIUM_KEY"],
)

Node

client.mjs

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic({
  baseURL: "https://proxium.tech",
  apiKey: process.env.PROXIUM_KEY,
});

curl

Terminal

curl https://proxium.tech/v1/messages \
  -H "x-api-key: $PROXIUM_KEY" \
  -H "Content-Type: application/json" \
  -d "{\"model\": \"$PROXIUM_MODEL\", \"max_tokens\": 256, \"messages\": [{\"role\": \"user\", \"content\": \"Hello\"}]}"

Send a message

A request is a list of messages and a limit on the length of the reply. Give model and max_tokens: without one of them, Proxium answers 400.

Python

message = client.messages.create(
    model=os.environ["PROXIUM_MODEL"],
    max_tokens=512,
    system="You answer in one sentence.",
    messages=[{"role": "user", "content": "What is a virtual key?"}],
)
print(message.content[0].text)

Node

const message = await client.messages.create({
  model: process.env.PROXIUM_MODEL,
  max_tokens: 512,
  system: "You answer in one sentence.",
  messages: [{ role: "user", content: "What is a virtual key?" }],
});
console.log(message.content[0].text);

The answer has the Anthropic message shape:

Field	Holds
content[0].text	The reply
model	The model that answered
usage	The token counts. Proxium prices the request from them

Stream the answer

With streaming, the reply arrives in small pieces while the model writes it, so your app can show it as it grows. To stream, use the stream helper of the SDK, or set stream to true. Proxium sends the pieces as Anthropic server-sent events.

Python

with client.messages.stream(
    model=os.environ["PROXIUM_MODEL"],
    max_tokens=512,
    messages=[{"role": "user", "content": "Write a haiku about queues."}],
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)
print()

Node

const stream = client.messages.stream({
  model: process.env.PROXIUM_MODEL,
  max_tokens: 512,
  messages: [{ role: "user", content: "Write a haiku about queues." }],
});
stream.on("text", (text) => process.stdout.write(text));
await stream.finalMessage();
process.stdout.write("\n");

curl

Terminal

curl -N https://proxium.tech/v1/messages \
  -H "x-api-key: $PROXIUM_KEY" \
  -H "Content-Type: application/json" \
  -d "{\"model\": \"$PROXIUM_MODEL\", \"max_tokens\": 512, \"stream\": true, \"messages\": [{\"role\": \"user\", \"content\": \"Write a haiku about queues.\"}]}"

Check which fields reach the model

Proxium passes on	Proxium drops
system, messages, tools, tool_choice, temperature, top_p, top_k, stop_sequences, stream, metadata.user_id	cache_control, container, mcp_servers, service_tier
Text, image, tool_use and tool_result blocks	thinking, redacted_thinking and document blocks

note

Proxium drops cache_control, so Anthropic prompt caching does not apply to a call through Proxium.

Handle errors

An error comes in the Anthropic error shape, and the SDK raises the class for its status. A body that is too large is the one exception: it gets a 413 with a plain-text body.

Error body

{
  "type": "error",
  "error": {
    "type": "authentication_error",
    "message": "invalid or revoked key"
  }
}

Status	error.type
400, 415, 422	invalid_request_error
401	authentication_error
402, 403	permission_error
404	not_found_error
413	request_too_large
429	rate_limit_error
529	overloaded_error
Any other status	api_error

A 429 has a retry-after header with the seconds to wait. The message names the limit, for example source budget exceeded for 'support-bot' (calls). Errors lists each reason.

Next steps

• Route requests to models: model ids, tiers and the fallback order.
• Track spend: name each app with x-proxium-source, and see what it costs.