Direct answer

Can Claude Code use non-Claude models?

A third-party gateway can technically translate Claude Code's Anthropic-shaped requests to a compatible non-Claude route, but Anthropic explicitly says it does not support routing Claude Code to non-Claude models through any gateway. Treat such use as a third-party compatibility experiment, not an Anthropic-supported configuration. Tools, MCP tool search, Remote Control, streaming, caching, and future Claude Code protocol changes need separate verification.

Last reviewed 2026-08-26 · Editorial review: KeepRouter Editorial

Start with Anthropic's support boundary

Anthropic's current gateway documentation says a gateway with a supported API format can connect to Claude Code. The same page says Anthropic does not endorse, maintain, or audit third-party gateway products and does not support routing Claude Code to non-Claude models through any gateway. A protocol translation can work for a tested request while remaining outside Anthropic's product-support scope.

Capability	Claude on a supported route	Claude through a third-party gateway	Non-Claude route through a gateway
Messages request	Supported	Gateway must track the current protocol	Third-party translation only
Tools and streaming	Supported for the selected Claude model	Gateway must preserve current fields and events	Test the exact model and translated route
MCP tool search	Normal product behavior	Non-Anthropic base URLs have extra forwarding requirements	Do not assume support
Remote Control	Product and account dependent	Non-Anthropic hosts have documented restrictions	Do not promise it
Anthropic support	Yes, within the selected service	Limited to the documented gateway boundary	No

KeepRouter setup

export ANTHROPIC_BASE_URL=https://keeprouter.com
export ANTHROPIC_AUTH_TOKEN=sk-kr-your-key
claude

The base URL, gateway credential, and model are separate controls. Anthropic notes that setting only ANTHROPIC_BASE_URL without a gateway credential can leave a saved claude.ai login active. Verify both variables and the selected model before assuming the serving or billing path changed.

Use a Claude model from the live catalog when you need the clearest Anthropic support boundary. If you deliberately select a non-Claude Messages-compatible route, record it as a third-party compatibility test and keep a direct rollback to Claude. The Claude Code guide covers current setup and failure diagnosis.

Compatibility is a test result

Claude Code sends Messages requests with content blocks, tools, streaming events, multi-turn state, evolving headers, and model-specific features. A gateway can preserve or translate supported fields. It cannot add a model capability, make tool decisions match Claude, or guarantee that a later Claude Code release behaves the same way.

Test a real repository task

Start a fresh session from the same commit, use the same prompt and allowed tools, and set a clear completion criterion. Record whether the task finished, which tools ran, human corrections, input and output tokens, latency, and charge. Test long history, tool-result continuation, errors, and any screenshot input your workflow uses.

Protect automated loops

Create a key limited to the evaluated coding models and set a spend boundary. Keep application or session step limits, avoid automatically replaying side-effecting commands, and distinguish authentication, credit, rate-limit, and upstream failures.

When to keep Claude

Keep a Claude model when its provider-specific behavior, support contract, or tested task performance is required. A non-Claude model is an additional evaluated option, not an automatic drop-in claim.

Frequently asked questions

Which environment variables are required?

Set ANTHROPIC_BASE_URL to the gateway and ANTHROPIC_AUTH_TOKEN to a scoped gateway key.

Does Anthropic support non-Claude models in Claude Code?

No. Anthropic's current gateway documentation says it does not support routing Claude Code to non-Claude models through any gateway.

Can third-party translation still work technically?

It can for a tested request path, but that does not change Anthropic's support boundary or prove tools, streaming, model features, and future versions will work.

Can I compare cost from a static article?

Use the live model pages and your own measured task usage because rates and token behavior can change.

Should I reuse my general production key?

Prefer a separate key restricted to evaluated coding models with an appropriate spend boundary.