magpie : the local gateway that connects 60+ coding agents to any model, from a single menu bar
🔎 An agent router, at last
If you use Claude Code in the morning, Codex in the afternoon, and Gemini CLI in the evening, you know the ritual: three config files, three API keys, three quota systems, and zero visibility into what it all actually costs. Each agent lives in its own bubble, with its provider locked in, and switching models for a given agent amounts to hacking together environment variables.
That's exactly the problem magpie solves — the open source project by yetone — the developer known to Neovim users — which has exploded on GitHub with over 15,000 stars and a spot in the trends for the week of October 18, 2026. The principle fits in one sentence: every agent's model, one place. A local gateway that installs in your menu bar and translates, in both directions, the protocols of all your coding agents.
The promise is bold: run Claude Code on Kimi, Codex on DeepSeek, or Gemini CLI on GLM-5.3, without touching a single line of code in the agents. And based on the feedback and the project's growth, it really works.
The Essentials
- magpie is a local gateway (127.0.0.1:3425) that speaks OpenAI Chat, OpenAI Responses, Anthropic Messages, and Gemini — and translates in both directions, including streaming, tool calls, and reasoning.
- ~55 supported agents: Claude Code, Codex, Gemini CLI, OpenCode, Cursor CLI, Zed, VS Code Chat, JetBrains Air, Crush, Devin, Cline, Droid, Qwen Code, Grok Build, Raycast… and Hermes Agent.
- 40+ providers: DeepSeek, Kimi, GLM, Qwen, OpenRouter, Ollama — plus your Claude, ChatGPT, GitHub Copilot, or Gemini Code Assist subscriptions turned into reusable providers.
- Automatic failover: when a quota runs out, magpie switches to the next model in your routing group, without any intervention.
- One-command installation: native binary under 15 MB, Docker image available on ghcr.io/yetone/magpie, TUI included.
Recommended Tools
| Tool | Main use | Price (October 2026) | Ideal for |
|---|---|---|---|
| magpie | Local multi-agent gateway | Free, open source | Multi-agent developers |
| Ollama | Local LLMs as a provider | Free | Open source models locally |
| OpenRouter | Model aggregator via a single key | Pay per use | Access to 100+ models |
| Claude Code | Terminal coding agent | Subscription or API | Agentic development |
If you're getting started with local agents, our guide on open source AI agents with Ollama and our comparison of the best LLMs to run locally pair well with magpie: the gateway accepts Ollama as a provider, so your local models become usable by any agent.
The problem it solves: agent fragmentation
Direct answer: before magpie, every coding agent was forcibly married to its provider, and changing spouses required configuration surgery.
Let's take a concrete case. Claude Code expects the Anthropic API. Codex expects OpenAI. Gemini CLI expects Google. If you wanted to test DeepSeek V4 Pro — which dominates the open source leaderboards with a score of 88 — under Claude Code, until now you had to hack around with environment variables, homegrown proxies, or often flaky community forks.
magpie flips the logic. The gateway sits between the agent and the model. The agent speaks its native protocol to magpie, and magpie translates to the provider of your choice. You write:
magpie provider add deepseek sk-...
magpie claude moonshot/kimi-k3
magpie gemini zhipu/glm-5.3
Three commands, and Claude Code runs on Kimi K2.6, Gemini CLI on GLM-5.3. The gateway even writes the agents' configuration for you: ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN for Claude Code, GOOGLE_GEMINI_BASE_URL for Gemini CLI, provider magpie for AtomCode. It's this level of polish that explains the project's virality.
How it works technically
A protocol translator, not a simple proxy
The technical difficulty isn't routing HTTP traffic — any reverse proxy can do that. It's correctly translating protocols that don't think the same way. Anthropic streaming doesn't have the same event format as OpenAI streaming. Tool calls have different schemas. Reasoning (the chains of thought exposed by some models) doesn't exist everywhere.
magpie handles all three: streaming, tool calls, and reasoning are translated in both directions. Concretely, Claude Code can trigger a tool call to a DeepSeek model without either side realizing they don't speak the same language. This is what distinguishes a serious gateway from a weekend hack.
The gateway lives with the app
magpie runs on 127.0.0.1:3425 and lives alongside the menu bar application. No separate daemon to supervise, no systemd service. The binary is under 15 MB, written in Go, installable via:
curl -fsSL https://usemagpie.ai/install.sh | sh
or go install if you prefer to compile. A TUI is available for those who live in the terminal.
Your subscriptions become providers
This is the most underrated feature of the project, and probably the one that will make publishers grind their teeth.
If you pay for a Claude or ChatGPT subscription, magpie can connect it via magpie accounts add codex and turn it into a provider. Same goes for GitHub Copilot, Gemini Code Assist and Grok. The result: your existing subscription powers any supported agent, not just the publisher's own.
The economic argument is crystal clear. Many developers pay for Claude Pro for Claude Code, ChatGPT Plus for Codex, and Gemini Code Assist for Gemini CLI — three subscriptions, each used at 30% of its quota. With magpie, a single well-used subscription can cover most of your agentic needs, with failover kicking in when the quota runs dry.
We're obviously in a gray area regarding the publishers' terms of service — something to evaluate before rolling this out at scale in a company.
Failover and routing groups: no more quota errors
The short answer: magpie lets you define model groups with order or rotate strategies, and automatically switches when a provider goes down or runs out of quota.
The typical use case: you route Claude Code to Claude Opus 4.7 (Adaptive), with Kimi K2.6 as fallback, then DeepSeek V4 Pro. Your Anthropic quota runs out at 5pm on a Thursday? The agent keeps running, magpie has already switched. You only notice when reading the cost logs.
Per-key cost tracking completes the picture: every request is tracked, provider by provider, so you can finally know whether that GPT-5.5 test on Codex cost you €2 or €40. For anyone who has ever received a surprise API bill at the end of the month, that alone is a selling point.
46 agents, including the less obvious ones
The official list catalogued on SkillsLLM turns heads: Claude Code, Codex, Gemini CLI, OpenCode, Cursor CLI, Zed, VS Code Chat, JetBrains Air, Crush, Devin, Cline, Droid, Qwen Code, Kimi Code, MiniMax Code, Grok Build, Raycast… and Hermes Agent.
Two notable points:
Grok Build is supported. xAI launched its first coding agent CLI recently, and the coding agents war is intensifying — magpie is already on board, which makes it an excellent vantage point for comparing agents against each other without changing your infrastructure.
Hermes Agent too. Emerging projects get integrated quickly, a sign of an active community and an extensible architecture.
On the model side, coverage is just as broad: DeepSeek V4, Kimi K2.6, GLM-5.1, Qwen3.6, MiniMax M2.7 on the open source side; GPT-5.5, Gemini 3 Pro Deep Think, Claude Opus 4.7 on the proprietary side. If you're hesitating about which model to plug behind your agents, our 2026 LLM comparison details the strengths of each.
npm plugins: extensibility as middleware
Another interesting architectural decision: magpie accepts npm plugins that run as middlewares in the gateway. Concretely, you can intercept and transform requests in transit — custom logging, redaction of sensitive data, caching, conditional routing.
It's the kind of mechanism that distinguishes a mature developer tool from a prototype. Teams working on proprietary code will be able, for example, to filter what goes out to cloud providers and force certain patterns to local models via Ollama. The data sovereignty question becomes manageable without giving up cloud agents.
What this changes concretely
For a solo developer, magpie removes a daily friction: testing a model under an agent becomes a one-liner instead of a config debugging session. Want to know if Gemini 3.5 Flash — the fast model that beats Opus 4.7 and GPT-5.5 on agentic benchmarks at 289 tokens/second — lives up to its promises on your codebase? Two commands and you have your answer, benchmarks or not.
For a team, the benefits are different: a single point of billing and traceability, shared routing groups, and the ability to standardize the fallback (for example, "never send code to non-self-hosted models") at the gateway level rather than on each developer's machine.
And for the open source ecosystem, it's an accelerator: models like DeepSeek V4 Pro or Kimi K2.6 become testable in agents they weren't designed for, which forces a head-to-head comparison with proprietary models. At AI-master.dev, we believe this is exactly the kind of neutral infrastructure the market was missing.
If your budget is tight, our feature on the 34 providers and 635 free LLM endpoints pairs well with magpie: the gateway can route to multiple free providers with failover, which makes the approach genuinely viable in production. And if you're looking to use Claude Code without paying for the subscription, our free Claude Code guide explores the options — magpie has become a central piece of it.
❌ Common Mistakes
Mistake 1: Believing that magpie replaces the provider
magpie does not provide any models. It routes and translates. You still need API keys (or subscriptions) from the providers. The gateway is an interconnection layer, not a supplier.
Mistake 2: Ignoring the terms of service of subscriptions
Sharing a Claude or ChatGPT subscription as a provider via magpie accounts add is technically trivial, but legally questionable. For personal use, it's your risk to assess. In a business context, have your legal department validate before hooking subscriptions up to agents they weren't intended for.
Mistake 3: Routing everything without configuring failover
The routing group is the killer feature, not the simple model switch. If you connect a single provider without a fallback, you're recreating the original problem — one provider outage and your agent is down. Always configure at least two models per group.
Mistake 4: Neglecting tool call compatibility
Not all models handle tool calls with the same reliability. An open source model that's excellent at code generation may fail at complex tool orchestration. Test the agent + model combination on your real use cases before switching over.
❓ Frequently Asked Questions
Is magpie really free?
Yes, the project is open source under the MIT license and the binary is free. You only pay for the models you consume through the providers — or nothing at all if you route to Ollama locally.
Does it slow agents down?
The gateway runs locally on 127.0.0.1:3425 and the added latency is negligible compared to the models' response times. Streaming is preserved, so the experience remains smooth.
Exactly how many agents are supported?
The official list counts around 55, including Claude Code, Codex, Gemini CLI, Cursor CLI, Zed, Devin, Cline, Grok Build, and Hermes Agent. Unlisted agents often work too, as long as they follow the OpenAI or Anthropic protocols.
Can you use it as a team?
Yes, via the Docker image published on ghcr.io/yetone/magpie, which can be deployed on a shared server. Per-key cost tracking makes it easy to split expenses among team members.
Who is behind the project?
yetone, a developer well known in the Neovim ecosystem, with an active community — more than 15,000 GitHub stars, and 4,282 stars as of October 3, 2026 on the repository, two weeks after it opened.
✅ Conclusion
magpie does what no agent vendor wanted to do: break the lock-in between agents and models, with technical execution that lives up to the promise. Install it, plug your favorite agent into the model you've been meaning to try for months, and set up a routing group with failover — it's the kind of tool you never uninstall. To dig deeper, explore our guide to the best autonomous AI agents and this week's AI news.