magpie is a menu bar app and local model gateway for command line coding agents, built so that Claude Code, Codex, Gemini CLI and OpenCode can run on whichever model provider you pay for. It is aimed at developers who keep several of those agents installed and several model subscriptions alive at once, and who are tired of each agent being wired to the company that shipped it. The project is written in Go, published under the MIT licence, and has gathered roughly 10,100 stars since the repository appeared in late September 2026.
What it does
Every coding agent arrives locked to one provider: Claude Code to Anthropic, Codex to OpenAI, Gemini CLI to Google. magpie collects the models you have access to into a single list in the menu bar, and a click repoints an agent at a different one. The examples the project itself gives are the obvious cross pairings: Claude Code on Kimi, Codex on DeepSeek, Gemini CLI on GLM, OpenCode on a ChatGPT plan.
Two further behaviours matter in daily use. magpie can fall back to another provider when one fails, and when an account's quota runs out it moves on to the next account you have configured instead of stopping. The README also links a separate community plugins repository, so the provider list is meant to be extended rather than fixed.
How it works
One local gateway does the serving. magpie runs on your own machine, the agents talk to it, and it forwards each request to whichever provider the selected model belongs to. That is the part that makes a single list possible: the agents never have to understand each other's APIs, only the gateway's.
The second half is deliberately small. Rather than wrapping or forking the agents, magpie edits the one key in each agent's own configuration file that decides where its requests go. The agent keeps its own config and its own behaviour, and switching models becomes a one line change that magpie writes for you.
Getting started
The README badges list macOS, Windows, Linux, Docker and Termux, while the GitHub topics name only macOS, so expect the desktop builds to be the furthest along and treat the rest as less proven. No hardware, memory or GPU requirement appears anywhere in the material, which fits what magpie does: the gateway routes requests, it does not run a model.
Cost is the point to be clear about. magpie itself is MIT licensed and free, and it ships no credits or tokens of its own. You bring your own API key or subscription plan for every provider you want in the list, and the quota fallback only helps if you hold more than one account in the first place. Nothing in the material describes running a local model behind the gateway, so plan on paid provider accounts.
Installing is more than a clone. The download sits behind the project's own site, usemagpie.ai, and the release badge points at a second repository, yetone/magpie-releases, whose Releases page carries the builds; the code in yetone/magpie is the Go source. Documentation starts at usemagpie.ai/docs/start, with a longer reference at docs/reference.md in the repository, and there is a Discord for questions. The material prints no single install command, so this entry will not invent one.
When to use it / when not
The catch is the one the project states in its own words: you still need your own key or plan. magpie moves your agents between models, it does not make the intelligence cheaper, and a router is only as useful as the quota you already bought. The second caveat is age. The repository was created in late September 2026 and was still being pushed to in October 2026, so this is a young project with a short track record and no long release history in the material.
Use it instead of hand editing environment variables and agent config files when you change provider more than once a day, or when you want an agent to keep working after one account hits its limit. Don't use it if you run a single agent on a single plan: the one key magpie would rewrite is a key you can set yourself in a minute, and a gateway in the middle is one more moving part to debug. Treat the fallback as a convenience rather than a guarantee, since the material says it can fall back, not that it always will.
Developers who have accumulated a drawer full of coding agents and a mixed bag of provider plans are the people this repository is built for, and the design is the reason to take it seriously: one local gateway plus one edited config key is a smaller, more inspectable idea than a fork of every agent. If you only ever use one agent with one subscription, there is nothing here you cannot do by hand. If you juggle Kimi, DeepSeek, GLM and a ChatGPT plan across four agents and keep running into limits, a menu bar list that repoints any of them, and quietly rolls to your next account when a quota is gone, is worth the install.