Use the Anthropic Python SDK with a custom Base URL.

What was actually verified on production?

Six probes were sent to https://monkeytheme.com/v1/messages with a deliberately invalid test key, on 2026-08-09. Every probe returned HTTP 401 with a structured JSON body and the x-request-id header in the response. The sub2api backend log shows ingress_reject_reason=invalid_api_key with latency_ms=0-3 on every request, meaning the relay gates on auth before any upstream call.

Install and configure the SDK

The Anthropic Python SDK is on PyPI as anthropic. The README documents ANTHROPIC_BASE_URL and ANTHROPIC_API_KEY environment variables; the constructor also accepts base_url= and api_key= directly.

python -m venv .venv
source .venv/bin/activate
pip install anthropic

export ANTHROPIC_BASE_URL="https://monkeytheme.com"
export ANTHROPIC_API_KEY="***"

Do not commit the key into source. os.environ is fine for the snippet below; for production prefer a secrets manager or ~/.zshrc-level shell export.

import os
import anthropic

client = anthropic.Anthropic(
    api_key=os.environ["ANTHROPIC_API_KEY"],
    base_url=os.environ["ANTHROPIC_BASE_URL"],
)

Verify the key before sending real prompts

The Anthropic SDK does not expose models.list(). To confirm a key is alive, send a single-token client.messages.create() call with max_tokens=1. If the key is bad you get a 401 with a request ID; if the key is good you get a 1-token Claude reply with a model field that confirms the upstream model name.

resp = client.messages.create(
    model="claude-3-5-sonnet-20241022",
    max_tokens=1,
    messages=[{"role": "user", "content": "Reply with OK"}],
)
print(resp.model, resp.stop_reason, resp.usage)

If resp.model prints a Claude model name visible to your account, the SDK wiring is correct. The Anthropic SDK README pins anthropic-version to 2023-06-01 on every request; you do not need to set it yourself.

How should errors be handled?

The SDK raises anthropic.APIStatusError for non-2xx responses. For connection and timeout failures it raises anthropic.APIConnectionError. Log the request ID from the exception; do not log the API key.

import anthropic

try:
    resp = client.messages.create(
        model="claude-3-5-sonnet-20241022",
        max_tokens=64,
        messages=[{"role": "user", "content": "hi"}],
    )
except anthropic.APIStatusError as exc:
    print("HTTP", exc.status_code, "request_id", exc.request_id)
    raise
except anthropic.APIConnectionError as exc:
    print("network/TLS:", repr(exc.__cause__))
    raise

A 401 with code=INVALID_API_KEY from MonkeyAI is the same code the OpenAI route returns. The relay does not distinguish between Anthropic and OpenAI keys at this layer; both are MonkeyAI keys, and they live in the same backend.

Streaming, tools and retries

Streaming on /v1/messages uses Server-Sent Events. The Anthropic SDK returns an iterator you can for-loop over. Retries on 429 and 5xx are on by default; set max_retries=0 while debugging so a transient 429 does not silently mask the real failure.

with client.messages.stream(
    model="claude-3-5-sonnet-20241022",
    max_tokens=256,
    messages=[{"role": "user", "content": "Summarise this in one sentence."}],
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)
    final = stream.get_final_message()

Tools work the same way they do against api.anthropic.com. If a tool_use block comes back, the SDK parses it into resp.content. If you instead see a 400 from upstream after a 200 from the relay, the upstream account likely rejected the tool schema, not MonkeyAI.

Common configuration mistakes

• Setting base_url to https://monkeytheme.com/v1 with the SDK on top. The SDK already adds /v1; the resulting path becomes /v1/v1/messages and nginx returns 404.
• Setting api_key to the OpenAI-style sk-*** value. The Anthropic SDK passes x-api-key; the relay accepts it on /v1/messages, but the same string is the MonkeyAI account key, so do not confuse it with Anthropic's own keys.
• Stripping anthropic-version manually. The SDK sets it; removing it does not help and breaks upstream compatibility checks.
• Calling client.models.list() on the Anthropic client. There is no such method on the Anthropic SDK; use requests or curl against GET /v1/models directly.
• Treating a 401 INVALID_API_KEY as evidence of upstream outage. The 401 came from the relay's auth gate, not from the upstream account. Check the API key first.

Sources

• Anthropic Python SDK README (pinned revision)
• Official Anthropic Python SDK repository
• Anthropic API reference: client SDKs
• MonkeyAI: OpenAI Python SDK custom Base URL
• MonkeyAI: Anthropic Node SDK custom baseURL

FAQ

Related guides