magpie: One Menu Bar to Rule Every AI Agent Model on Your Machine

What this project does

magpie is a tiny menu-bar utility that solves a very modern annoyance: you have a dozen AI coding agents installed on your machine — Claude Code, Codex, Gemini CLI, OpenCode, Cursor, Copilot CLI — and each one binds to a different model with its own config file. Swapping models means hunting through settings.json, config.toml, opencode.jsonc or config.yaml one by one. magpie pulls every agent and its current model into a single panel. Click a value, pick a model, done. In plain terms, it is the model-switching brain for all your local agents, so you never again dig through scattered config files just to point Claude Code at Kimi or Codex at DeepSeek.

Why it blew up recently

The repo was created on September 23 and crossed 600 stars within two days. The reason is timing: AI coding agents have exploded in number and each locks you into a specific model with rigid, separate configuration. magpie lands squarely on the "I want to swap models freely" pain point. On top of that, the author yetone has a real following in the AI-tooling community, and the README is clean and demo-driven, which helped it spread fast through developer feeds and group chats.

Technical highlights

• One small binary. The desktop build is under 15 MB because it uses Wails to call the system's native webview instead of bundling a browser engine; the terminal-only build is about 7 MB. It ships on macOS, Linux and Windows, so your whole team can use the same tool regardless of OS.
• Surgical config edits. When you change one model, magpie touches only that one key. Comments, key ordering and indentation in your settings.json, config.toml, opencode.jsonc or config.yaml survive untouched, and writes are atomic so a crash mid-write cannot corrupt the file. This matters more than it sounds: most config editors rewrite the whole file and silently reformat your carefully tuned settings.
• One gateway for every agent. magpie runs a local gateway that speaks OpenAI chat completions, OpenAI Responses and the Anthropic Messages API, then forwards to whichever vendor serves the model. Codex, Claude Code, OpenCode and the rest all point at http://127.0.0.1:3425/v1 and pick from one catalog; a translation layer smooths over the protocol differences so you do not have to remember which agent speaks which dialect.
• Three shapes. There is a menu-bar dropdown, a normal window opened with magpie, a terminal TUI via magpie tui, and a plain CLI, so you can use whichever fits your workflow whether you live in the terminal or on the desktop.

Who it is for

It is for developers who juggle multiple AI coding agents daily — say you write business logic with Claude Code, run tests with Codex and do research with Gemini CLI, and you are tired of editing each config separately. It is also perfect for anyone who wants to route agents to cheaper models like DeepSeek, Kimi or GLM but finds per-tool configuration tedious. If you already keep a spreadsheet of "which agent uses which model," magpie replaces that spreadsheet with a live panel.

Quick start

brew install magpie
magpie
magpie tui

How it compares

Against hand-editing each agent's config, magpie centralizes model switching into one panel and removes the back-and-forth. Unlike cloud-hosted model gateways, it is fully local with zero external dependencies, so your configuration never leaves the machine and there is no monthly bill. The only caveat is that it is brand new, so support for more obscure agents' config locations is still catching up via the community, but the mainstream agents most people actually use are already covered well.

A practical example: imagine you want Claude Code on a cheap long-context model for refactoring and Codex on a fast model for unit tests. Without magpie you edit two files; with magpie you click twice. That small friction removed dozens of times a day is exactly why people are starring it.

A few practical notes

One detail I really like is that magpie is not a cloud model router. It lives entirely on your machine, so no API key ever leaves your computer and there is no extra network hop adding latency. The local gateway at 127.0.0.1:3425 is the glue that lets an agent speaking only the Anthropic protocol talk to a model served behind the OpenAI protocol, and vice versa. If you have ever been annoyed that Cursor insists on its own model picker while Codex demands a different endpoint, magpie quietly removes that friction. The atomic, surgical config writes are a small but real luxury: I have lost settings to editors that rewrote an entire file and destroyed my hand-tuned formatting, and magpie's approach means that never happens here. Because it is so new, you may occasionally hit an agent whose config path is not yet recognized; the issue tracker is where to request support, and the community is adding them quickly. For the vast majority of daily users on the big three operating systems, though, it simply works out of the box, and the TUI means you can use it over SSH on a headless box too.