Large language model (LLM) agents can now reason over multi-step tasks and invoke external tools, yet they remain confined to software APIs and cloud services. No widely adopted protocol exists for these agents to discover and operate physical IoT devices. Today, connecting even one device class to an AI agent requires a bespoke stack of MQTT brokers, REST wrappers, or vendor SDKs, none of which expose the structured tool schemas that agentic frameworks expect.We present MCP-Edge, a framework that extends the Model Context Protocol (MCP) to edge and IoT environments. MCP-Edge makes three contributions. First, it defines a four-tier device taxonomy (constrained MCUs, smart IoT nodes, BLE-only wearables, and Linux-capable edge computers) and maps an appropriate MCP implementation strategy to each tier. Second, it introduces MCP-Lite, a lightweight server that runs on devices with as little as 512 KB of RAM, while even smaller devices (down to 2 KB SRAM) participate through gateway-proxied tool definitions. Both paths are supported by CBOR-based message encoding, schema caching, and offline request buffering for intermittently connected hardware. Third, it provides a gateway that bridges cloud-native MCP transports (SSE/HTTP) to constrained-device channels (UART, BLE, local Wi-Fi), presenting all downstream devices as standard MCP tool providers to the cloud agent.We evaluate MCP-Edge on nine platforms: Arduino UNO, Seeed Xiao RP2040, ESP32-C3, ESP32-C6, M5Stack, Waveshare T5-E1-Touch-AMOLED, Colmi smart ring, Raspberry Pi, and Arduino UNO Q. These span from 2 KB SRAM microcontrollers to quad-core edge computers with 4 GB RAM. Benchmarks show that MCP-Edge adds moderate latency overhead relative to raw protocol baselines while keeping memory usage within the Class 1 constrained-device budget defined by RFC 7228.