Claude Code
Point Claude Code at Primex instead of api.anthropic.com. Usage is billed
through your prx- key.
Setup
export ANTHROPIC_BASE_URL=https://gateway.primexos.com
export ANTHROPIC_API_KEY=prx-YOUR_KEY
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1
claude # billed through Primex
The base URL has no /v1 suffix. Claude Code appends /v1/messages itself.
Keep CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 in your shell when testing
with claude --bare too.
Supported models
Use any enabled Anthropic model returned by the live catalogue. Set the
model with /model inside Claude Code or the ANTHROPIC_MODEL environment variable.
Claude Sonnet 5.5 is currently coming soon. When it becomes available, integrations
must use tool_choice: {"type": "auto"} or {"type": "none"}; this model rejects
forced tool selection. To turn off up-front thinking, use
thinking: {"type": "between_tools"} with low, medium, or high effort. Keep thinking
blocks unchanged when continuing a conversation. See
Anthropic's Sonnet 5.5 migration notes.
What this does not cover
Auto-mode classifier billing
Claude Code's no-charge classifier feature uses real server-side safety review,
not a blanket exemption for classifier-labelled requests. It requires the
provider to receive safeguards and return usable safeguard_results; availability
also depends on the provider's rollout. Client-side classifier fallback calls
remain billed as ordinary token usage.
If Claude Code says your session is ineligible, auto mode still works. To suppress the notice while explicitly keeping the billed client-side fallback, set:
export CLAUDE_CODE_AUTO_MODE_SERVER=0
This does not make classifier calls free or disable auto-mode safety checks.
Use /status to inspect the Auto mode server setting. See
Anthropic's classifier billing guide.
Other model providers
Claude Code is hardwired to Claude model names, so it can't drive GPT / Grok / DeepSeek through Primex. For those, use Cursor / Windsurf / Codex or the OpenAI SDK against the OpenAI-compatible endpoint.