Price terms as published

Each line is the vendor’s own wording, split from the price capture checked on 2026-10-06. Nothing here is rounded, converted or re-worded — the full sentences stay in the review below. Lines that carry only a price continue the plan named on the line above, in the vendor’s own order.

2 price terms for Magpie
#Plan / termChecked
1Free (open source, MIT)2026-10-06
2bring your own model keys or subscriptions2026-10-06

The short version

Magpie is a menu-bar application (also a CLI and a TUI) that runs a local API gateway and rewrites the model configuration of whatever coding agents you have installed. The pitch on its own front page is four words long — “Every agent’s model. One place.” — and the mechanism behind it is simple: Codex speaks the OpenAI Responses protocol, Claude Code speaks Anthropic Messages, Gemini CLI speaks Google’s, and magpie sits between them on 127.0.0.1:3425 translating in both directions, streaming, tool calls and reasoning included, so any agent can be pointed at any vendor’s model.

It is free, MIT-licensed, written in Go, and has no pricing page: usemagpie.ai/pricing returns a 404 and the site’s own words for it are “Free · Open source”. What you still pay for is the model side — your own API keys, your local Ollama server, or the Claude / ChatGPT / Copilot subscriptions you already sign in with, which magpie reuses as providers.

The project is thirteen days old and moving: the repository went up on 2026-09-23, was last pushed on 2026-10-06, and its release repo tagged v0.1.1076 the same day — three releases in one day, which is the cadence of a project still in active rebuild, not of a settled tool.

What it actually does

One screen for every agent. The README claims “over 35 agents in one list”; the landing page names Claude Code, Codex, Gemini CLI, OpenCode, Pi, Goose, Cursor CLI, Copilot CLI and Crush. Click a model in the list, pick another, and magpie rewrites only that one key inside the agent’s own config file — comments, key ordering and indentation are preserved, and the write is atomic. Picking the agent’s native model puts back exactly what was there before.

One gateway for every API. The local endpoint accepts OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and Google Gemini, and translates between them on the way to whichever vendor serves the model. Anything with a base-URL setting can use it (OPENAI_BASE_URL, ANTHROPIC_BASE_URL, GOOGLE_GEMINI_BASE_URL), and magpie never reads API keys out of environment variables — keys live inside magpie instead.

Routing that keeps going. Several models from several providers can sit in one routing group; when one hits a rate limit or runs out of quota the group moves to the next option, with smart and failover modes and per-group priority, weight, context limits and cost caps.

The subscriptions you already pay for, shared across agents. Sign in to Claude Code and that login becomes a provider usable by OpenCode, Pi or Goose; sign in to Codex and GPT shows up in Claude Code’s picker; a Copilot seat serves any agent. Tokens are refreshed the way the agent itself refreshes them, and nothing is copied into a key vault.

Catalog, presets, usage. Model lists come from models.dev, so a model released that morning appears on the next refresh. There are twenty-seven provider presets written out in the README (Anthropic, OpenAI, Gemini, DeepSeek, Kimi, Zhipu GLM, MiniMax, StepFun, Qwen, Baidu Qianfan, Tencent Cloud, Huawei Cloud MaaS, Volcengine Ark, Mistral, Groq, xAI, OpenRouter, Together, Fireworks, SiliconFlow, NVIDIA NIM, ModelScope, Ollama, LM Studio), plus importers for CC Switch, Claude Code, Codex and Alma. The gateway records tokens, cache hits and cost per agent and per model.

Small, native, no Electron. Under 15 MB, built on the system webview, signed and notarised on macOS, with Windows and Linux builds from the same repository. Install from a terminal with curl -fsSL https://usemagpie.ai/install.sh | sh, or download the Mac build directly.

Pricing

Item Price What it covers
magpie Free Menu-bar app, gateway, CLI/TUI, routing groups, profiles — no paid tier, no pricing page (verified 404 on 2026-10-06)
Source MIT GitHub yetone/magpie, Go, 5,202 stars
Your models Your own cost API keys you paste, Ollama or LM Studio locally, or the Claude / ChatGPT / Copilot subscriptions you already pay for

There is no free-versus-pro decision to make here, because there is no pro. The honest framing is that magpie moves the cost line rather than removing it: the app is free, and the thing it routes is whatever you were already going to spend on models.

The numbers behind the download

Everything in this section was read from the GitHub API and the project’s own README and website on 2026-10-06:

  • 5,202 stars and 372 forks, created 2026-09-23 — roughly 400 stars a day since launch.
  • Last push 2026-10-06, latest tag v0.1.1077, with three releases shipped on 2026-10-06 alone.
  • MIT licence, Go for the bulk of the code, with Objective-C/Swift for macOS tray glue, JavaScript for the provider config schemas, and a little Python tooling.
  • 121 open items sit on the tracker. GitHub’s counter combines issues and pull requests; 18 of them are pull requests, so roughly 103 issues are open — which is what a thirteen-day-old project on a 0.1.x version line looks like, and a fair warning that the surface still changes.
  • Third-party coverage exists but is thin: a source tour at PyShine describes it as “a protocol-translation gateway, a quota-aware routing engine, subscription bridges that drive real local binaries, and config editors careful enough to be trusted with files other programs also write”, while Toolin.ai and opentools.ai have directory listings and no user reviews yet.

Who this is for

  • Developers running more than one coding agent. If Codex and Claude Code are both open all day, the plumbing that stops the model question being asked six times an hour is the point.
  • People who want a cheaper model behind an expensive agent. Codex on DeepSeek, Claude Code on Kimi, Gemini CLI on GLM — the README’s own examples.
  • Anyone sharing one Claude, ChatGPT or Copilot seat across several agents without pasting the same key into four config files.
  • Not for non-technical users. This is a developer tool that edits configuration files and speaks API protocols; there is no hosted version and no team plan.

What we deliberately didn’t test

We have not installed magpie, pointed an agent at it, or compared latency and token costs through the gateway against a direct connection. Every capability line above comes from the README, the official website and the GitHub API as captured on 2026-10-06 — an official-basis reading, not a hands-on review, and we do not score it.

Two caveats a reader should weigh without us having to test them: the project is thirteen days old with a 0.1.x version line, so the config-editing behaviour can still change under you; and the subscription-bridging feature depends on each agent’s own login refresh working the way it does today, which is the part most likely to be affected by a vendor changing how its sign-in tokens are handled.

Sources