Claudemix

claudemix

Run one Claude Code session that mixes models: a Claude model as the orchestrator, GPT subagents as executors. Your Claude subscription login is never proxied, stored, or re-signed; your OpenAI/Codex subscription powers the executor lanes.

Why this shape

Claude Code's API endpoint is process-global, so per-agent routing needs a gateway in front. Existing mixed-model setups put the Claude subscription login inside the proxy too, which re-originates your Claude traffic under a spoofed client identity. This project does not do that: Anthropic-bound requests pass through untouched with the session's own credentials, which is the same base-URL gateway pattern Anthropic documents for enterprise LLM gateways. Only gpt-* requests are re-authed, with a local key, to a CLIProxyAPI instance that holds only your OpenAI/Codex OAuth.

Quick start (agent setup)

Paste the raw contents of SETUP.md into Claude Code (or any capable coding agent) on your machine and let it do the whole thing: install the splitter, add the claudemix shell function, create the GPT subagent, and run the verification tests.

Manual setup is the same four steps; SETUP.md doubles as the documentation.

Prerequisites

- Claude Code with a subscription (OAuth) login
- CLIProxyAPI on 127.0.0.1:8317 with a completed Codex login and a local sk-* API key in its config
- Node 18+

Usage

Inside the session, delegate to GPT with the sol agent type (Agent tool subagent_type: sol, or agentType: 'sol' in Workflow scripts). The main model, and any subagent without a gpt model pinned, stays on Claude as normal.

The three gotchas that matter

1. Route subagents via agent definitions, not the inline model param. A model: gpt-... line in ~/.claude/agents/sol.md frontmatter works; passing the same string inline…