A physical button for Claude’s permission prompts

Ever had Claude wait twenty minutes for your “yes,” while you made coffee in the next room?

This is the small annoyance of agentic work: most tool calls go through, but every now and then Claude wants to do something that needs a human confirmation. Delete a folder. Push a branch. Run a script it hasn’t run before. The agent stops and waits.

If I’m at the laptop, fine — I tap and it continues. If I’m not, the agent just sits there. I’ve come back twenty minutes later thinking the work was running, only to find a permission dialog patiently blocking everything.

The point of an agent is that it does things while you’re elsewhere. So this kept bugging me.

What I built

I bought an M5Stack M5StickS3 — a tiny ESP32-S3 dev board with a 240×135 screen, a six-axis IMU, two buttons, and a speaker. It fits in the palm of my hand.

Anthropic ships a reference firmware for an older model (the M5StickC Plus) that pairs with Claude Desktop’s Hardware Buddy panel over Bluetooth LE. I asked Claude to write my own firmware from scratch, against their protocol spec, for the newer S3.

The repo is at p3ob7o/hwbuddy-notifier-S3.

What it feels like

When Claude Desktop hits a permission prompt — delete this folder? — three things happen on the M5 within a second or two:

• The screen lights up with an animated character and the tool name.
• A short chime plays. (Silent if the device is face-up, since I’m already looking at it.)
• A pill selector appears: Approve / Deny.

I tap the side button to switch between options, and the front button to confirm. If I’m holding it loose and want to deny quickly, I just shake it — a vigorous shake while a prompt is active equals Deny.

Bluetooth range, so I need to be within reach of the laptop. That’s fine for how I work; I’m rarely more than a room away from the machine.

What’s under the hood

Three things make this work:

• Anthropic’s BLE protocol: Nordic UART Service, newline-delimited JSON. The desktop sends heartbeat snapshots and permission prompts; the device replies with permission decisions (once or deny) and status acks.
• LE Secure Connections with bonding. A six-digit passkey shows on the M5 during pairing; macOS asks me to type it on the desktop side. Transcript snippets that flow over the link are AES-CCM encrypted.
• A character GIF and a melody, both swappable. The repo ships with neither, by design, because I used Totoro as a character and chime, but I don’t own the rights to it so I can’t distribute it. Bring your own. tools/encode_gif.py resizes and encodes a GIF to the device’s color format; src/melody.h is a flat list of (frequency, duration) tuples. I picked the sheet music from the song, asked ChatGPT to translate it into frequencies and durations, made a few edits.

So this:

Became this:

1. C5 523.25 Hz 0.375 s
2. A4 440.00 Hz 0.1875 s
3. F4 349.23 Hz 0.1875 s
4. Rest 0.1875 s
5. C5 523.25 Hz 0.375 s
6. Bb4 466.16 Hz 0.375 s
7. G4 392.00 Hz 0.5625 s
8. Rest 0.75 s
9. Bb4 466.16 Hz 0.375 s
10. G4 392.00 Hz 0.1875 s
11. E4 329.63 Hz 0.1875 s
12. Rest 0.1875 s
13. Bb4 466.16 Hz 0.375 s
14. A4 440.00 Hz 0.375 s
15. F4 349.23 Hz 0.5625 s
16. Rest 0.75 s

Why this matters to me

I like auto mode, where Claude only asks for permission when it matters, anywhere beyond hobby projects I don’t feel comfortable unleashing bypass permissions mode.

So this little hardware buddy, Totoro in my case, makes it easy and fun to give the occasional permissions, or deny once in a while when Claude is too reckless.

It’s also turns out to be fun to have a physical object that means “the agent wants something from you.” A chime in your pocket is more present than a notification badge on a laptop you’ve stepped away from.

And it was a nice weekend project.