yetone/magpie

What it does

magpie is a local model gateway and desktop app for people who use several coding agents and several model providers. Each agent normally keeps its model, base URL and keys in its own config file; magpie collects providers, subscriptions and local models in one place and rewrites the selected agent's config when you pick a model. The gateway listens on 127.0.0.1:3425 and translates between OpenAI Chat, OpenAI Responses, Anthropic Messages and Gemini formats, including streaming, tool calls and reasoning README lines 98–107.

Key capabilities from the documentation:

• Model switching: choose provider/model for any supported agent from a searchable list; the agent config is rewritten atomically and existing comments and formatting are kept. Profiles save and restore the whole set of agent choices at once README lines 110–130.
• Providers and subscriptions: presets cover Anthropic, OpenAI, Gemini, Vertex, Bedrock, DeepSeek, Kimi, GLM, Qwen, OpenRouter, Ollama, LM Studio and any OpenAI- or Anthropic-compatible URL. Subscription sign-ins (Claude, ChatGPT, Copilot, Gemini, Grok and others) can also be exposed as providers README lines 151–174.
• Failover and routing groups: groups combine several models or accounts and distribute requests with strategies such as smart, order, rotate, usage and pace README lines 175–210.
• Privacy masking (off by default): secrets and personal data are replaced with placeholders before a request leaves the machine and restored in the reply, with rules and optional local-model detection README lines 131–150.
• Usage tracking, plugins, sharing and an MCP-based search/image tool are described in the README.

The README lists a large set of supported agents, from Claude Code and Codex to Cursor CLI, Gemini CLI, OpenCode, Zed, Cline and many others README lines 80–97. It is aimed at users juggling more than one agent, provider or account; if one agent on its own vendor plan is enough, the README says you don't need it README lines 36–48.

Quick start

1. Install. Download the app from the website, or run the install script README lines 49–79:

   curl -fsSL https://usemagpie.ai/install.sh | sh
   

   Alternatives include go install github.com/yetone/magpie@latest and the Docker image Docker docs. On Linux the script installs the desktop app and installs WebKitGTK 4.1 first when needed; without a desktop it installs the terminal build. --gui or --tui chooses explicitly, and --proxy or --mirror handles firewalls.

2. Add a provider. Open magpie, go to Providers → Add provider, pick a preset and paste an API key, or sign in with a subscription.

3. Pick a model. On the Agents page, choose a model for each installed agent. Start a new session and it uses the new model.

The CLI can do the same:

magpie provider add deepseek sk-…
magpie claude deepseek/deepseek-v4-pro
magpie codex moonshot/kimi-k2.5
magpie group add "Opus anywhere" models=claude/claude-opus-5-5,copilot/claude-opus-5.5 routing=smart
magpie claude group/opus-anywhere
magpie quota

Any other tool that accepts a base URL can point at the gateway, e.g. OPENAI_BASE_URL=http://127.0.0.1:3425/v1 with OPENAI_API_KEY=magpie README lines 92–96.

Constraints and privacy

• The app runs on macOS, Windows and Linux, including Docker and Termux; the interface is available as a menu bar app, window, TUI, web UI and CLI.
• Subscription discovery reads agent sign-in files and the macOS Keychain. It can be disabled in Settings → Privacy → Subscription accounts, with magpie accounts discovery off, or per run via MAGPIE_NO_ACCOUNTS=1. API-key providers work without it.
• Privacy masking is off by default and must be enabled. It can block requests (503) if masking is on but cannot run, and it should not be used on values that the task needs to see in clear text; such values can be placed under Never mask.
• The app sends a once-a-day usage ping (random id, version, OS, agent/provider/model ids, partner interaction counts). It can be disabled in Settings → Privacy or with DO_NOT_TRACK=1. Partner links are clicked through usemagpie.ai/go/… redirects README lines 297–300.
• The project is MIT licensed. The repository does not accept pull requests; feedback goes to Discord.

HotRepo's take

The README presents a broad, concrete tool: a local multi-protocol gateway, GUI and CLI management, failover groups, masking and plugin middleware. It clearly targets power users with several agents and subscriptions rather than a single-agent workflow, and the configuration edits and gateway endpoint are documented in enough detail to try it. There are no HotRepo ranking, star-history or discussion records for this repository, so no trend observation can be drawn from site data; the practical evaluation should rest on the feature coverage and the disclosed network/privacy behavior described above.