Let your coding agent develop firmware on your real board

Agentic Hardware-in-the-Loop (Agentic HIL) connects the coding agent you already use to the board on your desk. The agent builds your firmware, flashes it, checks what the board actually does, and uses that feedback to fix failures. You review the code and the recorded hardware result.

Use your existing firmware project

Install once, restart your agent, and open the project you want to work on. Agentic HIL supplies the local hardware tools; your agent uses your project's build system and toolchain.

The development guide covers setup, the development loop, supported probes, and recorded evidence.

Install

Linux and macOS, in any shell:

curl -LsSf https://agentic-hil.github.io/install.sh | sh

Windows, in PowerShell:

irm https://agentic-hil.github.io/install.ps1 | iex

The installer registers the skill and local MCP server for the Claude Code, Codex, and opencode CLIs it finds. Restart your agent once after installation. Installation covers prerequisites, other install paths, and checking the installer before running it.

See the hardware loop

The recorded demo shows firmware being written, flashed, and checked on a real board. The STM32 starter includes firmware, hardware test plans, and a deliberate defect to find and fix on a Nucleo-F446RE.

• Flash and restart a board, with two recorded runs and their limits.
• Send commands and check answers over UART or CAN, with a failing expectation and its fix.
• Check a test plan without a board, and understand what requires a real hardware run.

What you can connect

Hardware support follows the debug probe and its backend: ST-Link through OpenOCD or STM32CubeProgrammer, and CMSIS-DAP through pyOCD. Your target toolchain and backend must support the board you connect. The Nucleo is the reference example. Support and backend prerequisites describe the current scope.

A configuration outside the firmware repository defines the devices and actions the agent may use. Hardware actions return structured results and logs that the agent and reviewer can inspect. The safety model explains that boundary.