Skip to main content

A framework for developers and AI agents

Most distributed applications begin with a simple task—publish a camera frame, request an action, expose a tool—and gradually accumulate transport-specific code. Reconnection, queueing, correlation IDs, broker conventions, thread ownership, and media handling begin to shape the application itself.

MAGPIE draws a boundary between what the application means and how bytes travel.

Separate intelligence from embodiment​

Intelligence and embodiment have different deployment needs. A model may run in a cloud GPU service, planning on an edge computer, and low-level control on a robot. Cameras, microphones, speakers, and actuators remain close to the physical system.

MAGPIE lets those locations change without rewriting the behavior:

Design principles​

Application code should not speak transport​

The application calls write, read, call, or respond. A transport implementation owns sockets, broker topics, signaling, correlation, background I/O, and shutdown.

Interoperability is a default, not a bridge​

Python, C++, and TypeScript share message envelopes, frame fields, default MessagePack serialization, JSON-RPC semantics, and MCP behavior. A mixed-language deployment should not need a translation service. Applications may supply another serializer when every endpoint on that path uses the same format.

The core should stay understandable​

Four primitives cover most communication:

  • StreamWriter publishes values or frames to topics.
  • StreamReader receives topic-filtered streams.
  • RpcRequester makes a request and waits for its correlated response.
  • RpcResponder receives requests and dispatches handlers.

Higher-level schemas, nodes, tools, and MCP integration build on those primitives rather than bypassing them.

Dependencies should follow features​

Applications should not install a media stack to send dictionaries. MQTT, WebRTC, video, audio, discovery, and MCP are optional capabilities. This keeps small services small and deployment intent explicit.

Operational tools are part of the framework​

A protocol is easier to adopt when developers can inspect it. MAGPIE includes CLI readers, writers, request clients, video/audio utilities, discovery, and SSH tunneling so systems can be tested before custom UI or orchestration exists.

What MAGPIE is—and is not​

MAGPIE is a reusable communication layer for robotics, edge systems, distributed services, browser clients, and AI agents. It is not a centralized runtime, workflow engine, device registry, or cloud platform. You retain control of process layout, infrastructure, authentication, and deployment.

That narrow boundary is deliberate: MAGPIE connects components without owning the system around them.