Switch from Anthropic Claude SDK to Meliai Inference

Meliai exposes a fully Anthropic-compatible endpoint at /v1/messages, accepting the exact request shape and returning the exact response shape that the Anthropic SDK expects. You migrate by setting two environment variables — ANTHROPIC_API_KEY and ANTHROPIC_BASE_URL — and your existing code continues to run without modification. Claude model names such as claude-sonnet-4 are mapped automatically to open-weight equivalents served from EU infrastructure.

Get a Meliai API key

Create a key at meliai.ai/account/api/keys. Your key will have the format sk-mel-<KEY>.

Keep your existing ANTHROPIC_API_KEY value if any other services in your stack still call Anthropic directly. You can hold both variables in your environment simultaneously.

Set your environment variables

Override the two values the Anthropic SDK reads to locate the API:

export ANTHROPIC_API_KEY="sk-mel-<YOUR_KEY>"
export ANTHROPIC_BASE_URL="https://api.meliai.ai"

The SDK picks these up automatically — no code changes needed at this stage. ANTHROPIC_BASE_URL must point to the root URL (without a /v1 suffix); the SDK appends the path itself.

Run your existing code

Your Anthropic SDK calls work unchanged. Here’s a minimal example to confirm the connection:

import os
from anthropic import Anthropic

client = Anthropic(
    api_key=os.environ["ANTHROPIC_API_KEY"],
    base_url="https://api.meliai.ai",
)

response = client.messages.create(
    model="claude-sonnet-4",
    max_tokens=256,
    messages=[{"role": "user", "content": "Name three Hanseatic cities."}],
)
print(response.content[0].text)

Streaming, tool use, vision inputs, and system prompts all work through the same /v1/messages endpoint with no modifications to your feature-specific code.

Understand model name mapping

Meliai does not serve Anthropic’s proprietary Claude weights. Instead, when you pass a Claude model name, Meliai maps it to the closest open-weight equivalent available on EU infrastructure. For example:

Anthropic model name	Meliai equivalent
`claude-sonnet-4`	Open-weight model of comparable capability
`claude-haiku-3-5`	Fast, lightweight open-weight model
`claude-opus-4`	High-capability open-weight model

You can also supply a Meliai-native model ID directly — browse the full mapping at meliai.ai/hub or retrieve it programmatically:

curl https://api.meliai.ai/v1/models \
  -H "Authorization: Bearer $ANTHROPIC_API_KEY"

The Anthropic SDK’s beta.messages.create interface (used for extended thinking, prompt caching, and other beta features) targets the same /v1/messages endpoint on Meliai. Pass the corresponding betas header values as you normally would; Meliai forwards them to the underlying model where supported.

Meliai also exposes POST /v1/messages/count_tokens, the Anthropic-compatible token-count preflight endpoint. Call it before a full completion to estimate token usage and cost without incurring inference charges.

What changes and what stays the same

Aspect	Before (Anthropic)	After (Meliai)
SDK	`anthropic` Python / JS	Same SDK, unchanged
Endpoint	`api.anthropic.com`	`api.meliai.ai`
API key format	`sk-ant-...`	`sk-mel-...`
Model weights	Proprietary Claude	Open-weight equivalents
Data residency	Anthropic (US)	EU-only, GDPR-compliant
Environmental telemetry	Not available	Per-response `environment_impact`
Response shape	Anthropic Messages	Identical Anthropic Messages shape

Overview

Models & Routing

Features

Privacy

Switch from Anthropic Claude SDK to Meliai Inference

What changes and what stays the same

​What changes and what stays the same

What changes and what stays the same