Magpie: The New GitHub App That Lets Codex Run DeepSeek and Claude Code Run Kimi

Daniel Carr, MSW

Sep 23

3 min read

Updated: Sep 24

Magpie appeared on GitHub on September 23, 2026 and has already reached more than 600 stars and 35 forks in roughly a day. The premise is unusually easy to understand: it lets you choose which AI model powers coding agents such as Codex, Claude Code, Gemini CLI, OpenCode, Cursor CLI, Goose, Copilot CLI, and several others from one application.

That means Codex can use DeepSeek, Claude Code can use Kimi, Gemini CLI can use GLM, and several coding agents can share models from OpenAI, Anthropic, Google, xAI, Mistral, Qwen, MiniMax, Groq, OpenRouter, Ollama, and other providers.

What Magpie actually does

Coding agents normally expect particular model providers and configuration formats. Codex understands OpenAI interfaces, while Claude Code is built around Anthropic's system, and other agents have their own configuration files.

Magpie sits between those agents and the model provider. It runs a local gateway on your computer that understands the different APIs and translates requests when necessary. You select the model you want, and Magpie adjusts the relevant configuration so the coding agent sends its work through Magpie instead.

The result is a model switcher for AI coding tools rather than another coding agent competing with them.

Codex on DeepSeek is the obvious attraction

One of the most useful possibilities is running an existing coding interface with a completely different model behind it. Someone who likes Codex as a coding environment but wants to experiment with DeepSeek can select a DeepSeek model inside Magpie. The same idea applies to Claude Code with Kimi or another compatible model.

Magpie also supports local model servers such as Ollama and LM Studio. That creates a path for using familiar coding agents while directing some work toward models running on your own computer.

It can also share subscriptions between agents

Magpie goes further than API keys. The developer says existing sign ins from Claude Code, Codex through ChatGPT, GitHub Copilot, and several other supported services can appear as model providers inside Magpie. For example, a model available through a supported subscription can then become available to another compatible agent through Magpie's local gateway. The implementation differs by service, since some providers allow API style access while others require Magpie to communicate through the provider's own installed command line application.

This is one of the more ambitious parts of the project, and it is also where provider rules and compatibility can become complicated. One current GitHub issue reports that some Copilot models requiring the Responses API are exposed by Magpie but fail when routed through the older Chat Completions path.

Why Magpie is getting attention so quickly

Magpie solves a problem that has become increasingly common as AI coding tools multiply. Developers can end up with Claude Code, Codex, Gemini CLI, Cursor, OpenCode, Copilot, and local models installed at the same time, each with separate settings and model choices.

Magpie reduces those separate configurations to one screen. It can also save profiles, so a user could keep one collection of expensive models for demanding work and another collection using cheaper or local models.

The desktop application is under 15 MB and uses the operating system's own web view instead of bundling a large browser runtime. Magpie supports macOS, Windows, and Linux, while a terminal version is also available.

Is Magpie worth watching?

More than 600 GitHub stars in roughly its first day is enough to make Magpie unusual, but the utility is more persuasive than the star count alone. AI coding tools are becoming increasingly interchangeable at the interface level while the models behind them change constantly.

Magpie is effectively betting that developers will want to choose the agent interface and the underlying model separately.

That could become increasingly useful if coding models continue improving at different speeds and prices. Instead of abandoning a preferred coding agent whenever another company releases a stronger model, Magpie attempts to make the model itself replaceable.

For software that was barely a day old when reviewed, Magpie already has working builds across the major desktop platforms, active development, dozens of forks, and a clear reason for existing. This is one of the new GitHub projects worth watching.