Magpie
yetone/magpieReviewed by AI
What it does
magpie is a small app plus a local gateway that decides which AI model each of your coding agents talks to, so Claude Code can run on Kimi or Codex on DeepSeek. You add provider keys, or the subscriptions your agents are already signed in to, and magpie rewrites each agent's own config and routes the requests through 127.0.0.1:3425. It can move to your next account when a quota runs out, show tokens and cost, mask secrets inside prompts, and run on a server with Docker. It is for people juggling several agents, providers or accounts.
How easy it is to install
B · installer plus one simple step · Windows
- Windows B
- macOS B
- Linux C
macOS and Windows are a B: each release has a ready .dmg and .exe (the README says Mac builds are signed and notarised and update themselves), and the one extra step is adding a provider key or signing in. Linux is a C, because the install script may first install WebKitGTK through your package manager with sudo, or fall back to the terminal build.
Numbers on GitHub
- Stars
- 26,956on Oct 31, 2026
- Downloads
- 114,40755,838 this week
- Growth
- +4,352stars in 5 days
Stars over time
Facts
- Runs on
- Windows, macOS, Linux
- You need to install
- Nothing extra
- To build from source
- Docker, Go
- Disk space
- 72 MB–215 MB · from the installer
- Accounts or keys
- At least one model provider key or signed-in subscription, OpenAI, Anthropic, Google Gemini
- Cost
- Free, API costs extra · The app is free (MIT). It routes to model providers and plans you pay for yourself; it can also point at local models through Ollama or LM Studio.
- License
- MIT (Permissive)
- Can replace
- OpenRouter, Portkey
- Last update
- Nov 1, 2026
Risk signals (8)
- AttentionVery new repository with many stars
26968 stars in 38 days
automatic check
- Attentionserves your subscription to other programs
README 'Subscriptions': 'Codex / ChatGPT: your ChatGPT plan's models, in every other agent', plus 'Add several accounts per subscription and magpie fails over between them' and a 'Daily warm-up' that 'starts each Claude and ChatGPT account's 5-hour window at the times you choose'. Each vendor's own terms decide whether a plan may be used this way and through several accounts.
AI review
- Attentionreads sign ins from your computer
README: 'magpie finds the subscriptions your agents are signed in to by reading their sign-ins (Claude Code's and Codex's files, the Keychain).' It can be stopped with 'Settings → Privacy → Subscription accounts' or 'magpie accounts discovery off'.
AI review
- Attentioninstall pipes remote script to shell
README quick start: 'curl -fsSL https://usemagpie.ai/install.sh | sh', and the small print says that on Linux without WebKitGTK 4.1 the script installs it 'with your package manager (sudo asks for your password)'.
AI review
- Attentionplugins see every request
README 'Plugins': plugins are installed from npm ('magpie plugin add opencode-gemini-auth') and 'run on Bun, downloaded the first time one needs it'. A plugin can also be 'gateway middleware: JavaScript that sees what every agent sends and gets back', able to rewrite or refuse a request.
AI review
- Attentiontelemetry on by default
README 'Privacy': 'Once a day a released magpie tells us it is in use: a random id, its version and system, and which agents, providers and models it is used with.' It is turned off in Settings → Privacy or with DO_NOT_TRACK=1. Partner links go through 'usemagpie.ai/go/…, which counts the click', and icons are fetched through usemagpie.ai/api/icon.
AI review
- Infopaid partners shown first
The README has a Sponsors row (APINEBULA, 零词元, Kimi, APIKEY.FAN, WandAPI, AIGoCode) and the privacy section counts, 'for each partner listed first in the add sheet, how many times a day it was shown, opened and added', so the provider list inside the app puts partners first.
AI review
- Infosecret masking is off until you turn it on
README: 'once you turn masking on in Settings → Privacy (it's off by default), keys, passwords and emails in a prompt are masked on your computer before the request leaves it'. The optional local detection model is 'a 1.6 GB download, on the CPU'.
AI review
Evidence is shown in its original language.
What Radar could not check
- Whether each listed agent and provider works in practice
- Whether a vendor blocks or closes an account used this way
- How well the masking rules catch secrets in real prompts, and what they miss
- What the npm plugins you add do with the requests they see
- What you spend across the providers and plans it routes to