wangsiyi7/claudex

Claudex lets you run the Claude Code command line interface while swapping out the underlying AI model for a Codex authenticated GPT model instead of Claude. It works by starting a small local proxy on your own machine that translates between the two systems, so the familiar Claude Code terminal and tools stay the same, but the reasoning underneath comes from the other model. No OpenAI API key is required, since it authenticates through Codex's own login instead. A key feature is separate control over how much reasoning effort different parts of a session use. The main session, where you are actively working, can be set to think harder, while helper agents that investigate a codebase in parallel can run at a lower, cheaper effort level. Claudex ships with named presets, such as economy, balanced, quality, and maximum, each combining a main effort level, an investigator effort level, how many investigator agents run, and how many run at once. These can be applied instantly or set individually through configuration commands. There is also a team feature that spins up several read-only investigator agents to look at a codebase in parallel, each reporting findings back to one main process that then combines everything into a final answer at a higher effort level. For security, the local proxy only listens on the machine itself, not the wider network, its admin interface is turned off, and it protects itself with a random access token. Login credentials are kept outside the project folder rather than being copied into it. The tool requires Node.js 20 or newer, along with Claude Code and the Codex command line tool already installed. It has been fully verified on Windows, with macOS and Linux support reported as still being tested by users. Claudex is not affiliated with OpenAI, Anthropic, Claude Code, or CLIProxyAPI, and is released under the MIT license.