MCP-Edge

Extend the Model Context Protocol (MCP) to edge and IoT devices.

MCP-Edge lets cloud LLM agents discover and invoke physical hardware through the same tool interface they already use for software APIs. It provides:

• a gateway that bridges cloud-native MCP transports (SSE / HTTP) to constrained device channels (UART, BLE, local Wi-Fi), presenting every downstream device as a standard MCP tool provider;
• MCP-Lite, a lightweight MCP server for devices with as little as ~512 KB of RAM;
• a four-tier device taxonomy (constrained MCUs, smart IoT nodes, BLE-only wearables, Linux-class edge computers) that maps an MCP strategy to each tier;
• protocol adaptations for constrained links: CBOR encoding, schema caching, connection pooling, and offline request buffering.

Status: alpha, under active development. The public API is unstable and will change. This repository is a reference implementation of the framework described in the MCP-Edge paper (IEEE Cloud Summit 2026). Performance figures reported in the paper are projected estimates, not measurements taken from this codebase. The full gateway path is exercised against simulated devices. A serial/UART transport has landed in the codebase — validated in software against an in-memory fake, but not yet exercised on physical hardware or paired with example firmware; BLE and Wi-Fi transports arrive in a later version.

Installation

Requires Python 3.10+.

pip install mcp-edge

For development, install from source:

git clone https://github.com/jemsbhai/mcp-edge
cd mcp-edge
pip install -e ".[dev]"

Quickstart

Run a demo gateway — an MCP server exposing two simulated devices over stdio:

mcp-edge run --demo

This serves a simulated sensor (read_temp, read_humidity) and a simulated ring (heart_rate), each tool namespaced by device (sensor-01/read_temp, ...). To drive it from an MCP client such as the MCP Inspector or Claude Desktop, configure a stdio server with command mcp-edge and arguments ["run", "--demo"]. The process logs to stderr and waits for a client on stdin.

Use the gateway as a library:

import asyncio

from mcp_edge.client import MCPLiteClient
from mcp_edge.devices import SimulatedDevice
from mcp_edge.gateway import Gateway
from mcp_edge.registry import DeviceRegistry
from mcp_edge.tiers import Tier
from mcp_edge.transports import LoopbackTransport


async def main() -> None:
    device = SimulatedDevice("sensor-01")
    device.add_tool("read_temp", lambda args: {"celsius": 21.5})

    transport = LoopbackTransport(device.handle)
    await transport.open()

    registry = DeviceRegistry()
    registry.register("sensor-01", MCPLiteClient(transport), Tier.SMART_NODE)

    gateway = Gateway(registry)
    print([tool["name"] for tool in await gateway.list_tools()])  # ['sensor-01/read_temp']
    print(await gateway.call_tool("sensor-01/read_temp", {}))     # {'celsius': 21.5}


asyncio.run(main())

Connect your own device over serial (experimental) — install the serial extra (pip install "mcp-edge[serial]") and point SerialTransport at the port your device is on. It exposes the same Transport interface, so it drops into the example above in place of LoopbackTransport:

from mcp_edge.transports import SerialTransport

# "COM3" on Windows, "/dev/ttyUSB0" on Linux/macOS
transport = SerialTransport("COM3", baudrate=115200)

Frames use a length-prefixed wire format — a 2-byte big-endian length followed by the CBOR payload — documented in transports/serial.py. Device firmware must frame its replies the same way; the exported encode_frame / decode_frame helpers make that straightforward. This path is not yet validated on physical hardware, and example firmware is still in progress.

Roadmap

• [x] v0.1 — gateway core, in-process (loopback) transport, protocol adaptations (CBOR, schema caching, connection pooling, offline buffering), device simulator, health monitor, CLI, hermetic CI
• [ ] v0.2 — real transports and Wokwi firmware-in-the-loop tests (Arduino / ESP32 / RP2040). The serial/UART transport (pyserial) has landed (software-tested); BLE and Wi-Fi/mDNS are still to come
• [ ] v0.2+ — Edge Impulse (inference as an MCP tool) and Arduino IoT Cloud (properties as MCP) integration examples; Renode / QEMU backends

Development

pip install -e ".[dev]"
pytest -q
ruff check .

License

MIT — see LICENSE.