Skip to main content
Use Auriko as your LLM backend for Claude Code and the Claude Agent SDK.

​
Prerequisites

​
Set up Claude Code

Add three environment variables to your shell profile (~/.zshrc or ~/.bashrc): 
export ANTHROPIC_BASE_URL="https://api.auriko.ai"
export ANTHROPIC_AUTH_TOKEN="ak_live_..."  # from auriko.ai/dashboard
export ANTHROPIC_API_KEY=""
Then claude works. Every Claude Code session routes through Auriko.
ANTHROPIC_API_KEY="" must be explicitly set to empty. If it contains a value, Claude Code uses it directly against Anthropic, bypassing the gateway.

​
Set up Claude Agent SDK

The Claude Agent SDK spawns the Claude Code CLI as a subprocess. Pass the environment variables through ClaudeAgentOptions.env: 
import os
from claude_agent_sdk import ClaudeAgentOptions, query

options = ClaudeAgentOptions(
    model="sonnet",
    system_prompt="You are a code review assistant.",
    allowed_tools=["Read", "Grep", "Glob"],
    env={
        "ANTHROPIC_BASE_URL": "https://api.auriko.ai",
        "ANTHROPIC_AUTH_TOKEN": os.environ["AURIKO_API_KEY"],
        "ANTHROPIC_API_KEY": "",
    },
)

async for message in query(prompt="Review main.py for bugs", options=options):
    print(message)
As defensive practice, pass setting_sources=[] in options to disable loading filesystem settings. This guarantees your env values take effect regardless of SDK version.

​
Use non-Claude models

Claude Code supports any model via Auriko’s routing: 
claude --model gpt-4o
claude --model deepseek-chat
claude --model gemini-2.5-flash
Switch models mid-session with /model gpt-4o, or override default aliases: 
export ANTHROPIC_DEFAULT_SONNET_MODEL="gpt-4o"
export ANTHROPIC_DEFAULT_OPUS_MODEL="claude-opus-4-6"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="claude-haiku-4-5"
Model names are the same canonical names you use with Auriko’s OpenAI endpoint.

​
Control costs and routing

Without Auriko, Claude Code talks directly to Anthropic. One provider, one price, no controls. With Auriko:
  • Cost visibility: see what Claude Code costs per session, per key, per team member
  • Budget controls: cap spend per API key so a runaway agentic loop doesn’t burn the budget
  • Provider routing: same Claude model available through Anthropic direct, Bedrock, and Vertex; Auriko picks the cheapest or fastest based on workspace config
  • Fallbacks: if Anthropic direct is rate-limited or down, automatically fail over to Bedrock or Vertex
  • Data policies: enforce zero data retention (ZDR) or no-training across all Claude Code usage
  • Unified billing: Claude Code usage in the same dashboard as all other LLM usage
  • Non-Claude models: claude --model gpt-4o routes through Auriko to OpenAI

​
Use the Anthropic SDK directly

Any application built on the Anthropic SDK can use Auriko as a drop-in backend: 
import os
import anthropic

client = anthropic.Anthropic(
    base_url="https://api.auriko.ai",
    api_key=os.environ["AURIKO_API_KEY"],
)

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello"}],
)

​
Configure routing

Pass routing options through extra_body: 
response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello"}],
    extra_body={
        "gateway": {
            "models": ["claude-sonnet-4-6", "claude-opus-4-6"],
            "routing": {"optimize": "cost"},
        },
    },
)
For Claude Code, configure routing at the workspace level via the Auriko dashboard. The CLI doesn’t support per-request routing options.

​
Check compatibility

  • All Claude Code features work through Auriko: tool use (Bash, Read, Write, Edit), MCP servers, streaming, extended thinking, and prompt caching
  • Agentic features (file operations, bash execution, MCP) run locally in the CLI; only LLM inference calls route through Auriko
  • Non-Claude models work for simple tasks but degrade on complex multi-file refactors (this is inherent to Claude Code, not Auriko-specific):
CapabilityClaudeGPT-4oDeepSeek/GeminiLocal models
Tool callingExcellentGoodSimple calls workUnreliable
Edit accuracy~98%~85-90%~60-80%Worst
Long sessionsStableDegradesDegrades fasterDegrades fastest
  • Fast mode’s org eligibility check is hardcoded to https://api.anthropic.com and doesn’t respect ANTHROPIC_BASE_URL. Fast mode is silently disabled when routing through a gateway.