Guide

No Anthropic Account Needed: Claude-Flavored Tooling on Every Model

AtmoRouter serves an Anthropic-compatible Messages endpoint for its whole catalogue — Claude Code and Anthropic SDK users can point their existing tools at any model with a base-URL swap.

Oct 6, 20264 min readAtmoRouter Team

A lot of the best tooling in 2026 is wired for the Anthropic Messages API: Claude Code, a generation of agent frameworks, and plenty of internal scripts. Until now that choice dictated your backend. It doesn't have to.

AtmoRouter serves an Anthropic-compatible Messages endpoint for its entire catalogue. Tools that speak the Anthropic protocol run unchanged — you swap a base URL and key, and pick any model.

Claude Code

Claude Code reads its endpoint from the environment:

bashexport ANTHROPIC_BASE_URL="https://atmorouter.dev"
export ANTHROPIC_AUTH_TOKEN="ar-xxxxxxxxxxxx"
claude

The ANTHROPIC_AUTH_TOKEN is your AtmoRouter dashboard key. Model selection inside Claude Code then maps onto AtmoRouter model ids, and the rate card applies per model.

Anthropic SDK (Python)

pythonfrom anthropic import Anthropic

client = Anthropic(
    base_url="https://atmorouter.dev",
    api_key="ar-xxxxxxxxxxxx",
)

msg = client.messages.create(
    model="zai/glm-5.3",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello in one word."}],
)
print(msg.content[0].text)

Note there is no /v1 — the Anthropic SDK appends its own path, same as always. Streaming works identically (client.messages.stream).

Why this is worth caring about

  1. 1.Tool lock-in is model lock-in. If your agent framework only speaks one protocol, your model choices are whatever that protocol's vendor sells. A compatible endpoint breaks that coupling in both directions.
  2. 2.Cost comparison per turn, not per ecosystem. With Claude-flavored and OpenAI-flavored tools drawing from the same balance, you can move a workload between model families and watch the actual number move — no separate accounts to reconcile.
  3. 3.Failed requests don't bill. The never-bill-failed-requests policy applies on both endpoints, and agent tooling retries constantly.

A note on migration

Both endpoints are served from the same key and balance, so moving an existing tool is reversible — swap the base URL back and you are where you started. Try it first on a tool you can afford to break for an evening, then move the rest.

Mentioned in this post

Start building today

Create an account, grab a key, and make your first call in minutes.