M5 OnboardSAFE
Official, Anthropic-managed directory of high quality Claude Code Plugins.
Overview
Official, Anthropic-managed directory of high quality Claude Code Plugins.
76137dccbc88OBSERVED · 2026-10-02Host compatibility
What the documentation claims. We have not run a compatibility test.
| Host | Status | Notes |
|---|---|---|
| claude-code | mentioned | |
| claude-desktop | mentioned | |
| cursor | mentioned |
What it tells the agent
The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.
---
name: m5-onboard
description: End-to-end onboarding for a freshly-plugged-in M5Stack ESP32 device (Cardputer, Cardputer-Adv, Core, CoreS3, Stick) — detect on USB, flash UIFlow 2.0 firmware, and install the Claude Buddy MicroPython app bundle. Use whenever the user plugs in or wants to flash/provision/reset an M5Stack or ESP32 board, or says "m5-onboard go".
---
# M5Stack Onboarding
This skill automates the full cold-start workflow for an M5Stack ESP32 device: detect on USB, identify model, flash UIFlow 2.0, and push a MicroPython app bundle onto `/flash/` so the device boots into user software. The apps we ship (Claude Buddy, Snake, Hello) talk over BLE or USB. The workflow runs on macOS, Linux, and Windows; the skill was developed against an M5Stack Basic v2.6 (CH9102 bridge, ESP32-D0WDQ6-V3, 16 MB flash) and generalized to cover the rest of the Core family, with the Cardputer-Adv (ESP32-S3, native USB) as the current default target.
## Where the scripts live
This skill ships as part of the `cwc-makers` plugin for reference, but the executable scripts and the `buddy/` app bundle live in a local clone of https://github.com/moremas/build-with-claude (the `/maker-setup` command creates this clone). Run every `scripts/*.py` invocation below from inside that clone's `onboard/` directory so `--apps buddy` resolves to the sibling `buddy/device/` payload.
## When to use
Use this when a user plugs in an M5Stack device and wants it provisioned. The decision tree:
- **Fresh/unknown device** → run `onboard.py --apps buddy` end-to-end (detect → identify → flash → install apps). This is the default path.
- **Already-flashed device, user just wants apps installed/refreshed** → run `install_apps.py --src buddy` (or any `--src <path>` to a directory of `.py` files).
- **Flashed device, something feels broken** → run `smoke_test.py` (I2C + LCD + speaker + button check).
- **User wants to know what's on the bus / what the device can do** → `smoke_test.py`.
If multiple devices are plugged in, ask which port to target — don't guess. If the user is provisioning a device they previously worked with (e.g. "same thing as last time" or "another Buddy"), default to `--apps buddy` unless they say otherwise.
### Which variant to assume
The rig this skill lives on provisions **Cardputer-Adv** boards overwhelmingly, so `onboard.py` now defaults to `--variant cardputer-adv`. In practice that means:
- If the user says nothing about the model, go with the default. They're almost certainly holding a Cardputer-Adv.
- If the user says "Cardputer" (no "Adv"), ask — the two models share a form factor but take different firmware images, and flashing the wrong one boot-loops the device.
- If the user names any other board ("Core2", "CoreS3", "Basic", "Fire"), pass the matching `--variant` explicitly — the default won't apply.
- The chip is ESP32-S3 either way, and `detect.py` won't be able to tell Cardputer from Cardputer-Adv before UIFlow is flashed (same native USB-JTAG VID, no pre-flash I2C probe). So this is a user-intent question, not a hardware-fingerprint one.
## The workflow
The main orchestrator is `scripts/onboard.py`. It drives the sub-scripts in order and handles the handoffs between them (waiting for reboots, capturing MAC, reporting progress). Prefer calling it directly over stitching the sub-scripts yourself unless the user asks for a partial run.
The default provisioning command (fresh Cardputer-Adv, install the buddy bundle):
```
python3 scripts/onboard.py --apps buddy
```
**How to invoke this from Claude Code's Bash tool.** Do NOT call `onboard.py` as a foreground Bash command. The Bash tool captures output and does not stream it back to the assistant until the command exits — and this command runs 2–3 minutes. That silence looks identical to a hang, and the assistant will usually give up before the button-dance prompt ever reaches the user. Instead, always run with `run_in_background: true`, `tee` to a log file, and then use the Monitor tool (or periodic `tail` via Read) to surface stage banners, heartbeats, and prompts to the user in real time. `2>&1` is not the fix — all progress already writes to stderr, which a terminal shows fine. The fix is streaming semantics, not redirection. The pattern that works:
```
# Launch (background, tee log):
python3 scripts/onboard.py --apps buddy 2>&1 | tee /tmp/m5-onboard.log
# Monitor (surfaces key events without drowning in byte-progress spam):
tail -f /tmp/m5-onboard.log | grep -E --line-buffered \
"^====|heartbeat|Heads up|Enter download mode|download mode!|rebooted into UIFlow|Manual reset|DONE|ERROR|Error|Traceback|FAIL|failed|No USB|not detected|Attempt [0-9]|Device already in download|Download mode port|Post-flash port|Waiting for device"
```
### Relaying physical steps to the user (REQUIRED)
The flash stage **cannot proceed without a manual button press** on native-USB boards — there is no software path. When the monitored log shows `Enter download mode` (or the script appears to wait at the FLASH stage), you MUST stop and tell the user to do the following on the **back of the Cardputer**, in your own words, before continuing:
1. Press and **hold** the **G0** button
2. While still holding G0, briefly press and release the **RST** button
3. Keep holding G0 for about one more second, then release it
4. The screen should go fully dark — that means download mode is active
If the device reboots into UIFlow instead of going dark, tell the user G0 was released too early and to try again holding it longer. Do not move on, retry the script, or attempt a software workaround until the user confirms the screen is dark — the flash will not start otherwise. The same applies to any later `Manual reset` prompt: relay the physical step and wait for the user.
Users running `onboard.py` directly in their own terminal (not via Claude Code) will see all output live — no changes needed there.
If `--port` is omitted, `detect.py` picks the most likely candTrust audit
SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.
| Layer | What it checks | Result |
|---|---|---|
| L0 | Provenance & inventory | PASS |
| L1 | Static analysis of the code | NA |
| L2 | Instruction surface (what it tells the agent) | PASS |
| L3 | Class-specific surface | PASS |
| L4 | Behavioural (sandbox) | SKIPPED |
What the source does
- Filesystem
- none-observed
- Network
- none-observed
- Shell
- none-observed
- Dependencies
- pinned
- Secrets in source
- none-found
Findings (0)
No findings outside the package's declared scope.
Gates applied: no_behavioural_pass.
76137dccbc88full audit observations/trust-audit/skill/anthropics__m5-onboard.json · Report an issue / request a re-scanAudit history
Every audit this skill has had.
| Date | Source | Verdict | Grade | Score | Change |
|---|---|---|---|---|---|
| 2026-10-02 | 76137dccbc88 | SAFE | B | 89 | first audit |
Questions
What does the M5 Onboard skill do?
Official, Anthropic-managed directory of high quality Claude Code Plugins.
Is M5 Onboard safe to install?
The audit found nothing in the source that contradicts what it says it does, and graded it B (89/100). Grade A is held back for packages that have also passed a sandboxed behavioural run, which is why a clean skill reads B.
What can M5 Onboard access on my machine?
The audit observed no filesystem, network or shell use at all in its source.
Which assistants does M5 Onboard work with?
Its documentation mentions claude-code, claude-desktop and cursor. That is what the text claims, not a compatibility test we ran.
How current is this page?
The grade is for one exact copy of the source (76137dccbc88), read on 2026-10-02. The repository is watched, and a new audit runs when it changes — this is the first audit.