magpie

magpie is a multi-model API gateway + coding agent model switcher that runs locally (GitHub yetone/magpie, MIT-licensed, written in Go). It solves a specific pain point: Codex only speaks the OpenAI Responses protocol, Claude Code only speaks Anthropic Messages, and Gemini CLI speaks the Google format—magpie starts a local gateway at 127.0.0.1:3425 that accepts all three protocols and translates between them bidirectionally (including streaming, tool calls, and reasoning), letting any agent call any vendor's models.

Core Capabilities

• Protocol-Translating Gateway: Exposes standard paths such as /v1/chat/completions, /v1/responses, /v1/messages, /v1/messages/count_tokens, /v1beta/models/{model}:generateContent, and /v1/models; any client supporting a custom base URL can connect (OPENAI_BASE_URL / ANTHROPIC_BASE_URL / GOOGLE_GEMINI_BASE_URL, gateway key can be any value)
• Surgical Agent Config Rewriting: Changes only the key the user modified, preserving comments, key order, and indentation with atomic writes; covers ~/.claude/settings.json, ~/.codex/config.toml, opencode, goose, and more; restores precisely when switching back to native models
• Subscription Reuse: Uses logged-in Claude Code (Pro/Max/Team), Codex (ChatGPT Plus/Pro/Business), Copilot, Grok, Cursor, and Devin sessions as providers, whose models appear with claude/…, codex/…, copilot/… prefixes in other agents' model selectors, no key copying needed
• Vendor Presets and Dynamic Model Lists: About 20 built-in presets (Anthropic, OpenAI, Gemini, DeepSeek, Kimi, GLM, MiniMax, Qwen, Mistral, Groq, xAI, OpenRouter, Together, Fireworks, SiliconFlow, AiHubMix, 302.AI, Ollama, LM Studio, and more); with a key it pulls model lists from vendors in real time, supplemented by the models.dev directory for names and reasoning levels
• Routing Groups and Smart Routing: Merges multiple accounts/keys into one group/<id> with intent-based routing (a small-model classifier decides), sequential, round-robin, and least-used-first strategies, plus rate-limit awareness (per-vendor Retry-After), balance exhaustion, and failover
• Profiles Config Snapshots: Saves each agent's settings as named profiles (like "Budget," "Focus") and switches them wholesale in one click
• Multi-Entry and Lightweight: Menu bar desktop app (Wails + system WebView, not Electron), TUI (magpie tui), and CLI; supports macOS / Windows / Linux; desktop builds under 15 MB, terminal builds about 7 MB
• Secondary Capabilities: magpie://import import links (parameters sit in the URL fragment, never sent to the server by the browser), usage stats (tokens / cache hits / cost), self-update, and encrypted backups

Use Cases Suited to local developers who run two or more AI coding agents at once (Claude Code + Codex + Gemini CLI, etc.) and want unified model switching plus reuse of existing subscriptions, especially macOS users. Users with a single agent, or enterprise teams highly sensitive to compliance and credential security, will gain little.

Name Disambiguation (Important): This entry is yetone/magpie (usemagpie.ai), not Blinue/Magpie (the Windows screen magnifier), nor magpie.im, magpie.report, AMD Magpie, or various same-named Magpie projects in Postgres / data tooling.

Risk Warnings: (1) It reads other agents' credential files and reuses subscription logins across agents, and "warms up" accounts per reset windows; the README itself advises "use an account you can afford to lose," carrying a risk of violating vendor terms and getting banned; (2) the site's /terms and /privacy are both 404—no terms of service or privacy policy; (3) the project is extremely early (repo created 2026-09-23, version 0.1.x, iterating from v0.1.218 to v0.1.223 within a day); (4) single-maintainer (author yetone with 401 commits, the other 11 contributors with 1–8 each), no company entity or support commitments; (5) curl | sh install with silent auto-updates, and Windows/Linux builds are unsigned. On the plus side: it is genuinely complete MIT Go source, the gateway listens on loopback only, keys are stored locally with 0600 permissions, the README mentions no telemetry, and community reception is positive (1,242 stars / 69 forks).