Skip to main content
Claude Code is Anthropic’s terminal coding agent. Point it at OrcaRouter with environment variables.
Protocol: Anthropic Messages
  • Base URL: https://api.orcarouter.ai — without /v1
  • Claude Code appends /v1/messages itself; adding /v1 would produce /v1/v1/messages and fail

​
Install

npm install -g @anthropic-ai/claude-code
On Windows, Claude Code requires Git Bash. Install Git for Windows, then set CLAUDE_CODE_GIT_BASH_PATH to your bash.exe (e.g. C:\Program Files\Git\bin\bash.exe).

​
Configure

export ANTHROPIC_BASE_URL="https://api.orcarouter.ai"     # no /v1
export ANTHROPIC_AUTH_TOKEN="sk-orca-..."
export ANTHROPIC_MODEL="orcarouter/auto"                  # or e.g. anthropic/claude-opus-4.8
export ANTHROPIC_SMALL_FAST_MODEL="anthropic/claude-haiku-4.5"

​
Run

claude -p "Reply with exactly: OK"
You should see OK. Then run claude for an interactive session and use /model to switch models.

​
Notes

  • ANTHROPIC_BASE_URL must not include /v1.
  • Model names use the vendor/model format; orcarouter/auto is the simplest choice.
  • On Windows, set CLAUDE_CODE_GIT_BASH_PATH to your Git Bash bash.exe.
VS Code extension: Anthropic’s official Claude Code extension just runs the CLI inside the editor — it shares the same configuration. Set the environment variables above (or your ~/.claude/settings.json) and the extension picks them up; there’s no separate in-editor endpoint setting.

​
Enable thinking / reasoning

Extended thinking is on by default (adaptive) — a reasoning model thinks automatically, no extra config needed. To tune it, lower the effort level with the /effort command or via /model, disable thinking in /config, or set MAX_THINKING_TOKENS to cap the budget on fixed-budget models (adaptive models ignore a nonzero budget — use effort levels there).