Zephyr 4.5: MCP Server, QUIC, and FIDO2 on MCUs Zephyr 4.5, released October 28, ships a built-in Model Context Protocol server that turns every enabled sensor on a networked device into a tool an AI coding agent can call, registering a sensor_read_ tool for each Devicetree binding declaring class: sensor. The release also adds QUIC/HTTP3, SSH-2 server and client support, FIDO2 over USB and BLE, RTP (RFC 3550), IPv4 NAT with connection tracking, Infineon TriCore and Qualcomm Hexagon architecture support, 178 new boards, and SPDX 3.0 software bills of materials for EU Cyber Resilience Act compliance. Zephyr reports 21% higher TCP throughput and 12% higher UDP throughput in zperf loopback benchmarks, and documents an agent that read console output from a running sensor driver, diagnosed an interrupt-handler timing bug, applied a fix, and verified it from hardware output without human involvement. Zephyr 4.5 landed on October 28, and for most of its 10-year history, a new RTOS release meant new board support and driver fixes. This one is different. The standout feature is a built-in MCP server that lets AI coding agents read physical sensors directly from the device — a first in the RTOS world. The release also adds QUIC/HTTP3, SSH remote access, FIDO2 security keys over BLE, GStreamer-style multimedia pipelines, two new architectures, and SPDX 3.0 software bills of materials for EU Cyber Resilience Act compliance. MCP Server: AI Agents Can Now Read Your Hardware The most interesting addition is not a network protocol or a new chip target. It is the Model Context Protocol server that ships as a first-class Zephyr service. Any Zephyr device with network connectivity can now run as an MCP server, and when it does, every enabled sensor on the board becomes a tool that an AI coding agent can call. The implementation leans on Devicetree. When a sensor binding declares class: sensor , the MCP server automatically registers a sensor read