Use Claude Code with Any Model via OpenRouter

Posted on Nov 4 • Originally published at adrian-alvarez.com

I've been using Claude Code for a while, and at some point a question stuck: what if I don't want Claude as the model behind Claude Code?

It turns out the tool is more flexible than it looks. Claude Code reads a set of environment variables that control where its API requests go and which model each role uses. The important one is ANTHROPIC_BASE_URL: it points Claude Code at any endpoint that speaks the Anthropic Messages API, not only Anthropic's own.

Anthropic documents this for LLM gateways, and OpenRouter ships an Anthropic-compatible endpoint built for exactly this case.

Instead of talking to Anthropic directly, Claude Code can talk to:

https://openrouter.ai/api Enter fullscreen mode Exit fullscreen mode OpenRouter is one API in front of hundreds of models from dozens of providers, with a single key and a single bill. Claude Code keeps speaking its native protocol, and OpenRouter translates the request for whichever model you picked. No local proxy in between.

So you keep the Claude Code interface, its tools, the agent loop and the terminal experience, and you swap what sits underneath.

One caveat before going further. Claude Code is tuned for Anthropic's API semantics: tool calls, streaming, extended thinking. A non-Anthropic model may handle some of that worse, or not at all. OpenRouter itself recommends Anthropic's first-party provider for full compatibility. Treat other models as an experiment, not a drop-in.

export ANTHROPIC_BASE_URL = "https://openrouter.ai/api" export ANTHROPIC_AUTH_TOKEN = " $OPENROUTER_API_KEY " export ANTHROPIC_API_KEY = "" Enter fullscreen mode Exit fullscreen mode ANTHROPIC_BASE_URL sends every request to OpenRouter. ANTHROPIC_AUTH_TOKEN authenticates with your OpenRouter key. ANTHROPIC_API_KEY must be explicitly empty. If it holds a value, Claude Code can fall back to its normal Anthropic auth. The OpenRouter integration guide calls this out too. Then the model overrides. Claude Code uses model aliases internally ( opus, sonnet, haiku, fable ) plus a separate default for subagents. Each has its own variable, documented here:

export ANTHROPIC_DEFAULT_FABLE_MODEL = " $MODEL " export ANTHROPIC_DEFAULT_OPUS_MODEL = " $MODEL " export ANTHROPIC_DEFAULT_SONNET_MODEL = " $MODEL " export ANTHROPIC_DEFAULT_HAIKU_MODEL = " $MODEL " export CLAUDE_CODE_SUBAGENT_MODEL = " $MODEL " Enter fullscreen mode Exit fullscreen mode Set all of them. If you only change the main model, background tasks and subagents keep asking for a Claude model by alias, and you end up running two models without noticing.

Exporting eight variables by hand every time gets old, so I wrapped it in a command.

orclaude <vendor/model-slug> [claude args...] Enter fullscreen mode Exit fullscreen mode For example:

orclaude anthropic/claude-sonnet-5.5 orclaude openai/gpt-6.1-sol Enter fullscreen mode Exit fullscreen mode The script:

#!/usr/bin/env bash # Run Claude Code against any OpenRouter model: # orclaude <vendor/model-slug> [claude args...]: " ${ OPENROUTER_API_KEY:?Set OPENROUTER_API_KEY in your shell profile } " MODEL = " ${ 1:?Usage: orclaude <vendor/model-slug> [claude args...] } " shift export ANTHROPIC_BASE_URL = "https://openrouter.ai/api" export ANTHROPIC_AUTH_TOKEN = " $OPENROUTER_API_KEY " export ANTHROPIC_API_KEY = "" export ANTHROPIC_DEFAULT_FABLE_MODEL = " $MODEL " export ANTHROPIC_DEFAULT_OPUS_MODEL = " $MODEL " export ANTHROPIC_DEFAULT_SONNET_MODEL = " $MODEL " export ANTHROPIC_DEFAULT_HAIKU_MODEL = " $MODEL " export CLAUDE_CODE_SUBAGENT_MODEL = " $MODEL " exec claude --model " $MODEL " " $@ " Enter fullscreen mode Exit fullscreen mode The variables only live in that process. Your normal claude session stays on your usual account, and orclaude is the OpenRouter one. No proxy, no Docker container, no fork of Claude Code.

To confirm it took, run /status inside the session. It should show the auth token as the source and https://openrouter.ai/api as the base URL.

On Windows, add a wrapper to your PowerShell profile ( notepad $PROFILE ) so orclaude behaves like any other command:

function orclaude { & " $HOME \Scripts\orclaude.ps1" @ args } Enter fullscreen mode Exit fullscreen mode Since the model is only an argument, the PowerShell version can also keep a shortlist of favorites and let you pick one by number:

# Run Claude Code against any OpenRouter model: # orclaude <vendor/model-slug | list number> [claude args...] $Models = @( "anthropic/claude-sonnet-5.5" "openai/gpt-6.1-sol" "google/gemini-3.8-flash" "deepseek/deepseek-v4.1-flash" ) if ( $args. Count -eq 0 -or $args [ 0 ] -in @( "-help", "--help", "-h" )) { Write-Host "" Write-Host "Available models:" -ForegroundColor Cyan for ( $i = 0; $i -lt $Models. Count; $i ++ ) { Write-Host " [ $( $i + 1 ) ] $( $Models [ $i ] ) " } Write-Host "" exit 0 } if ( -not $ env: OPENROUTER_API_KEY ) { Write-Error "Set OPENROUTER_API_KEY in your environment" exit 1 } $Model = [ string ] $args [ 0 ] if ( $Model -match '^\d+$' ) { $Index = [ int ] $Model if ( $Index -lt 1 -or $Index -gt $Models.

Count ) { Write-Error "No model # $Index. Run orclaude -help to see the list." exit 1 } $Model = $Models [ $Index - 1 ] } $ClaudeArgs = @( $args | Select-Object -Skip 1 ) $ env: ANTHROPIC_BASE_URL = "https://openrouter.ai/api" $ env: ANTHROPIC_AUTH_TOKEN = $ env: OPENROUTER_API_KEY $ env: ANTHROPIC_API_KEY = "" $ env: ANTHROPIC_DEFAULT_FABLE_MODEL = $Model $ env: ANTHROPIC_DEFAULT_OPUS_MODEL = $Model $ env: ANTHROPIC_DEFAULT_SONNET_MODEL = $Model $ env: ANTHROPIC_DEFAULT_HAIKU_MODEL = $Model $ env: CLAUDE_CODE_SUBAGENT_MODEL = $Model & claude --model $Model @ ClaudeArgs exit $LASTEXITCODE Enter fullscreen mode Exit fullscreen mode Two details in there are easy to get wrong:

Available models: anthropic/claude-sonnet-5.5 openai/gpt-6.1-sol google/gemini-3.8-flash deepseek/deepseek-v4.1-flash Enter fullscreen mode Exit fullscreen mode and orclaude 3 starts a session on the third one. A number outside the list fails with a message instead of launching Claude Code with an empty model.

For macOS or Linux, the Bash script above is the whole setup. Save it as orclaude somewhere in your $PATH and make it executable:

chmod +x orclaude Enter fullscreen mode Exit fullscreen mode Then:

orclaude <vendor/model-slug> Enter fullscreen mode Exit fullscreen mode If you'd rather have every session go through OpenRouter, put the variables in ~/.zshrc or ~/.bashrc instead. That is the persistent setup OpenRouter recommends. Claude Code does not read.env files, so the shell profile is the place.

The script is short. The interesting part is the architecture it exposes.

Claude Code is the harness: the tools, permissions, the agent loop, the terminal UI. The model is a separate layer, and the harness does not need to come from the same vendor as the model.

Claude Code → Anthropic-compatible gateway → model

With OpenRouter as that middle layer you get one API key, one bill, one catalog, provider routing with failover, and usage tracking in one dashboard. Switching models is an argument, not a migration.

You keep the Claude Code workflow you already know, and you can still find out what a different model does inside it.