magpie: every coding agent's model, from one menu bar

If you run Claude Code and Codex side by side, you know the ritual: to point Codex at DeepSeek for an afternoon you open ~/.codex/config.toml, edit the provider table, fix the model id, and hope the TOML still parses; switching Claude Code to Kimi means finding ~/.claude/settings.json and starting over. magpie puts one menu bar panel over all of it: every agent on your machine and the model it is set to, listed in one screen — click a value, pick a model. yetone open-sourced it on September 23, 2026, and it had 1,584 GitHub stars as of September 29; the WeChat article that brought it around calls it a new rival to CC Switch. What separates it from a plain config switcher is a local gateway: every agent points at one endpoint, and the API differences get translated there.

When it runs, the panel looks like this (the README’s TUI demo, same keys):

  ◉ magpie

  ▸ Claude Code   claude-fable-5-1[1m]                        ~/.claude/settings.json
    Codex         gpt-6-astra   effort medium
    Gemini CLI    gemini-3.1-pro
    OpenCode      anthropic/claude-sonnet-5   small anthropic/claude-haiku-4-5
    MiMo Code     anthropic/claude-sonnet-5
    Pi            openrouter/z-ai/glm-5.2:batch
    Goose         anthropic/claude-sonnet-5
    Cursor        auto
    Copilot CLI   claude-fable-5

  ↑↓ agent  ·  ←→ field  ·  ↵ change  ·  s save profile  ·  p profiles  ·  q quit

Core features

• One panel, every agent: more than 20 agents are covered — Claude Code, Claude Desktop, Codex, Gemini CLI, OpenCode, Cursor CLI, Copilot CLI, Crush, Devin among them; only the ones you have installed show up. magpie touches just the one key you change, so comments, ordering and indentation in your settings.json or config.toml survive, and writes are atomic.
• One gateway for all of them: magpie serves OpenAI Chat Completions, OpenAI Responses and Anthropic Messages on 127.0.0.1:3425, translating between the three — streaming and tool calls included. Agents stop holding vendor keys and URLs; every model is spelled provider/model.
• Subscriptions, shared: a signed-in Claude Code, Codex (ChatGPT), Copilot or Devin login shows up as a provider in every other agent’s picker. Claude subscriptions run through the real claude binary, with the caller’s tools bridged into the same turn over MCP.
• Routing groups: bundle several models into a group/<id> and route requests smart (subscriptions with soonest-renewing quota first), by order, rotate, or least-used; a conversation can also stick to the account that answered it.
• A provider takes one field: DeepSeek, Kimi, GLM, MiniMax, Qwen, OpenRouter, Ollama and 20-odd other presets need only a key. Model lists come live from each vendor, with models.dev filling in the names — a model released this morning is in the picker after the next refresh.
• Profiles: magpie save work snapshots every agent’s current settings; magpie use work puts them all back in one move.

Typical use cases

• Run Codex on DeepSeek today and keep Claude Code on Kimi for good, without memorizing either config format — two clicks in the menu bar, restart Codex, done.
• Feed a whole bench of agents from one Claude subscription: OpenCode, Pi and Goose all borrow the same login and model catalog through the gateway.
• Stretch several accounts: when two providers serve the same model, their keys become one group, and smart routing picks the subscription whose quota renews soonest.
• A box you only reach over SSH: magpie tui runs in the terminal, and magpie web serves the same screen in a browser (WSL and servers included).

Quick start

Download the app for macOS, Windows or Linux from usemagpie.ai, or install from a terminal:

curl -fsSL https://usemagpie.ai/install.sh | sh

Open magpie, add a provider (magpie provider add deepseek sk-…), then hand any agent a model:

magpie claude deepseek/deepseek-chat
magpie codex gpt-5.6-sol

Agents read their config at startup: a running session keeps its model, and the next one picks up the change.

Summary

magpie is for anyone running more than one coding agent and swapping models or vendors often, and for anyone whose Claude or ChatGPT quota is idling inside one CLI. If you run a single agent and never touch models, skip it — the value is in the many. If your pain goes beyond model switching to dispatching tasks across machines and approving from your phone, Paseo is the heavier layer; for models and subscriptions alone, magpie stays light.

A few honest catches: sharing a Claude subscription requires Claude Code installed and signed in on the same machine; Codex reads its model list at startup, so a switch needs a restart; the app sends one anonymous PostHog event a day (Settings → Privacy turns it off, or DO_NOT_TRACK=1); and wiring in an Antigravity account risks Google suspending it — the README itself suggests using “an account you can afford to lose”. MIT-licensed and written in Go; the macOS build is signed and notarized, the Windows and Linux builds are not, so SmartScreen will grumble once.