Magpie is the missing config layer that lets AI coding agents share any provider

The proliferation of local AI coding agents has created a genuinely tedious configuration problem: each agent stores its model choice in a different file format, in a different location, with different field names. Switching from Claude Code to DeepSeek for a project, or pointing Goose at a Kimi model instead of Anthropic's, means hunting down config files and editing them by hand. Magpie solves exactly this and nothing more.

The core mechanic is a unified list of every installed agent and its current model, editable from a single screen. The app surfaces as a menu bar icon, a normal window, a TUI, and a plain CLI — all showing the same table. Change a value, and magpie writes only that key back to the agent's native config file, leaving comments, ordering, and indentation untouched. Writes are atomic. That surgical approach matters: agents like OpenCode use JSONC with comments that a naive JSON writer would destroy.

The more interesting piece is the local gateway running on port 3425. It speaks OpenAI chat completions, OpenAI Responses, Anthropic Messages, and the Gemini API simultaneously, translating between them as needed — streaming and tool calls included. Every agent points at 127.0.0.1:3425 and uses provider/model notation; magpie handles the routing. This means you can point Claude Code at a DeepSeek model, or Gemini CLI at a Kimi model, without touching any agent's own provider configuration.

The subscription-sharing feature is the sharpest edge here. If you are signed into Claude Code, Codex, or Copilot, magpie exposes those logins as providers that any other agent can use. No key copying, no credential duplication. The Claude case requires a workaround worth understanding: Anthropic classifies requests from other agents as third-party traffic even when the OAuth token is genuine, so magpie drives the actual local claude binary for every Claude subscription request, bridging the caller's tools over MCP. It is a real constraint acknowledged directly, not papered over.

The binary is under 15 MB with the desktop app — it uses the system webview through Wails rather than bundling a browser runtime — and 7 MB for the terminal-only build. Model lists are live: magpie queries each vendor's API for its current catalog and refreshes from models.dev in the background, so a model released today appears in the picker on the next sync rather than waiting for a binary update.

Profiles let you snapshot every agent's settings under a name and restore all of them at once. Backup and restore are encrypted with AES-256-GCM, key derived via PBKDF2-SHA256, and handle the cross-machine case where some agents may not be installed on the target.

The scope is deliberately narrow. Magpie does not manage agent sessions, does not intercept prompts, does not add observability beyond a gateway debug log. It is a config manager with a routing layer, and it does those two things with unusual care for the files it touches.