Magpie turns the multi-agent config mess into a single switchboard

The problem magpie solves is embarrassingly mundane and genuinely annoying: you have six AI coding agents installed, each with its own config file format, and switching all of them from one model to another means hunting through settings.json, config.toml, opencode.jsonc, and whatever else before you can actually start working. Magpie collapses that into a single panel — menu bar, window, TUI, or plain CLI — where every installed agent and its current model are visible at once.

The config-editing behavior is worth pausing on. Magpie doesn't rewrite your files; it touches only the specific key it changes. Comments, indentation, ordering — all preserved. Writes are atomic. This matters because these config files are often hand-tuned, and a tool that silently reformats them would be unusable in practice.

The more interesting piece is the local gateway. Magpie runs an OpenAI/Anthropic/Gemini-compatible proxy on 127.0.0.1:3425 and translates between APIs — streaming, tool calls, reasoning levels included. The practical payoff: sign into Claude Code once, and every other agent on your machine can use that subscription's models through the gateway, with no key to copy and no credentials stored in magpie itself. The same applies to Codex (ChatGPT login) and Copilot (GitHub login). Magpie reads the agent's own credential store, refreshes tokens the way the agent would, and writes rotated tokens back where the agent expects them.

The Claude subscription path is the most architecturally unusual part. Anthropic classifies requests from other agents as third-party traffic even when the OAuth looks like Claude Code. Magpie's solution: it drives the actual local claude binary for every Claude subscription call, bridges the caller's tools in over MCP, and keeps the generated harness out of Anthropic's system-prompt classifier. This requires Claude Code to be installed and signed in — it's a real dependency, not a footnote.

The binary is under 15 MB for the desktop build (system webview via Wails, nothing bundled) and 7 MB for the terminal-only build. The model list is live: magpie fetches what each vendor actually serves, with the models.dev catalog filling gaps, refreshing in the background. A model released this morning shows up on the next sync.

Routing groups let you treat several models — across providers or accounts — as one target, with configurable strategies: smart (quota-aware, soonest-renewing first), round-robin, least-used, or ordered fallback. Session stickiness is configurable too, which matters for cache efficiency.

The telemetry disclosure is clean: one daily ping to PostHog with a random install ID, version, and platform. No keys, no models, no prompts. Opt out with DO_NOT_TRACK=1. Source builds never send it at all.

Twenty-plus agents supported, from Claude Code and Codex to Goose, Cursor CLI, Devin, and several less-familiar names. The breadth suggests this is tracking the actual fragmented landscape rather than the popular-press version of it.