axiometa-cli 1.0.4

Prompt to hardware, in your terminal: detect a board, chat to build, deploy, watch serial.

axiometa-cli

Prompt → hardware, in your terminal. The CLI is a thin client of the same backend Studio's browser talks to: codegen stays in the cloud (Claude + the catalog + your auth/limits), and only the local half — provision the toolchain, deploy, read the serial channel — runs on the machine the board is plugged into. Same brain as Studio; the terminal (or a Pi) is the body instead of the browser.

It is one experience whether the board speaks Arduino or MicroPython — the board's facts decide the deploy verb (deploy = compile+flash on Arduino; run/install on MicroPython).

Install

Raspberry Pi / Linux / macOS (from a checkout):

./cli/install.sh
axiometa provision        # one-time: arduino-cli + ESP32 core + libs (+ mpremote)
axiometa                  # detect a board and start building

Windows / any OS (pip):

pip install -e cli
axiometa provision
axiometa

Config (env, all optional):

AXIOMETA_BACKEND=https://studio.axiometa.io   # or your local backend
AXIOMETA_TOKEN=...                            # your auth (carried to the backend)
AXIOMETA_KNOWLEDGE=/path/to/backend/knowledge/boards   # auto-found next to a repo checkout

The experience (axiometa)

Typing axiometa with no verb runs the guided loop:

1. Detect — lists connected boards, naming the silicon from its USB identity (ESP32-S3 vs RP2040).
2. Target — confirms the board + language, picks the port, writes axiometa.json.
3. Toolchain — if arduino-cli/mpremote is missing, offers axiometa provision.
4. Chat — describe what to build; Axie streams its reply, shows what it's reading, and writes the project files. Refine in follow-ups ("make it blink faster").
5. Deploy — when the device program is (re)generated, offers to flash it (Arduino) or run/install it (MicroPython), streaming the toolchain output.
6. Monitor — after a successful deploy, streams the serial channel so you see the device talk.

Mid-chat commands: /flash /deploy /run /install /monitor /serve /board /help /exit.

Verbs (scripting / power use)

axiometa provision               # install the local toolchain (one-time)
axiometa devices                 # list connected boards (with USB identity)
axiometa board pi-hat            # set the target (writes axiometa.json)
axiometa prompt "blink the RGB LED on port 1 red"   # one-shot: generate + write files
axiometa flash --firmware mp.uf2 # RP2040 over SWD: lay MicroPython on the chip (Pi HAT)
axiometa run                     # MicroPython: test in RAM (ephemeral)
axiometa install                 # MicroPython: persist to flash
axiometa deploy                  # Arduino: compile + flash
axiometa monitor                 # stream serial to the terminal
axiometa serve                   # serve browser/*.html on 127.0.0.1 (ssh -L hint when headless)

The Pi HAT (RP2040 flashed over SWD, talked-to over UART)

The Pi HAT's RP2040 has no USB — the Raspberry Pi is both its programmer and its host, over the 40-pin header. So the Pi HAT path differs from a USB board, and the CLI handles it:

• axiometa flash --firmware <mp.uf2> lays MicroPython onto the chip over SWD (OpenOCD bit-banging Pi GPIO24=SWDIO / GPIO25=SWCLK / GPIO23=RUN). A .uf2 is auto-converted to the .bin OpenOCD needs; .elf/.bin also work. This is one-time, and the brick-recovery path.
• run / install / monitor talk over the Pi's UART at /dev/serial0 (RP2040 GP0/GP1 ↔ Pi GPIO15/14), not USB. The CLI ships board_pins.py + boot.py with each deploy; boot.py puts the REPL on UART0 so mpremote can reach the board over /dev/serial0.

One-time Pi setup (handled by axiometa provision, or do it yourself):

sudo apt install openocd                 # SWD flashing
sudo raspi-config  -> Serial Port:  login shell over serial = NO, serial hardware = YES   # frees /dev/serial0
sudo -E axiometa flash --firmware mp.uf2 # SWD needs GPIO/root; -E keeps your env

Notes: the flashed firmware must expose the UART REPL (freeze boot.py into the build, or use a UART-REPL MicroPython) for first contact. bcm2835gpio (the default OpenOCD driver) is Pi 0-4 only; on a Pi 5 set AXIOMETA_OPENOCD_DRIVER=linuxgpiod. Override the whole OpenOCD interface with AXIOMETA_OPENOCD_CFG=/path/to.cfg if your wiring differs.

Where things live

A project is a folder with an axiometa.json manifest (board + substrate + modules), the device program (sketch.ino / main.py), any browser/*.html surfaces, and board_pins.py. cd into a folder and run axiometa — it is project-rooted, like git.

Headless vs. desktop

The CLI detects where a visual surface can render and tells the backend (the capability_profile): on a Pi with a display it serves the surface to a local browser; headless-over-SSH it prints an ssh -L line to open it from your own machine; a bare terminal degrades a surface to a serial view.

Status

The agentic loop is board-agnostic and built: detect → chat (SSE codegen) → flash/deploy → monitor, plus a cross-platform provision. The Pi HAT's SWD flash (OpenOCD) and UART (/dev/serial0) deploy/monitor path is wired and unit-tested (uf2→bin, config generation, dispatch) but not yet validated on real hardware — exercise it on a connected Pi + RP2040. The one external dependency is a MicroPython firmware that exposes the UART REPL.

The native host-linux runtime (the "whole Pi" brain: code using the host's own camera/files/GPIO) is not wired here — it is approval-gated and waits on the trust model in backend/knowledge/runtimes/host-linux/contract.md.