How to use Claude to build a physical approval gate
Claude built Nightshift’s device-control software while its owner tested an approval board, with timed grants, revocation, rollback, and honest stand-ins.
devpost.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Implement this device-command control plane under the attached interface contract. Keep model proposals behind deterministic policy, time-limited physical approvals, snapshots, health checks, and rollback. Add FREE-WILi streamed-button and USB-reconnect handling for my hardware tests. Label simulated executors clearly and leave architecture, firmware flashing, and approval decisions to me.
Reach people exploring AI agents
See placements →What happened
Nightshift demonstrated a physical approval and revocation surface around gated device commands, with explicit rollback paths and documented hardware limitations.
Reported by Aneesh Grover — DevpostHow the agent did it
- Claude implementation sessions built the gate, executors, bridge, and dashboard under a shared interface contract.
- Grover brought up the FREE-WILi and Pixel hardware and corrected firmware and button-stream integration.
- The prototype enforced time-limited button approvals, revocation, snapshots, health checks, and rollback around agent commands.
Where a person stepped in
Grover designed policy and architecture, brought up and tested hardware by hand, and used physical approvals for risky actions; containers were clearly labeled stand-ins.
Aneesh Grover’s Nightshift prototype put a physical approval surface in front of agents acting on devices. Parallel Claude Code sessions implemented components under a written interface contract, while Grover handled architecture, hardware bring-up, and testing himself.
The system paired a FREE-WILi board with a device control plane, including a Pixel 9 Pro over adb. Model-proposed commands passed through a policy gate. Low-risk actions took snapshots and health checks; risky actions requested a time-limited grant. Buttons could approve, deny, or undo, while a double shake revoked grants.
The hardware integration had concrete failures. The shipped workshop firmware caused connection problems, button polling failed, and USB ports changed on reconnection. The implementation used streamed button events and reconnection handling after the owner restored stock firmware. Sound remained disabled because tones could stick.
Grover also replaced unreliable CLI model fallback with a purpose-built Gemini tool loop that recorded the selected model. The result was a working hackathon control layer with physical authorization and explicit container stand-ins. It was not evidence of fleet-scale savings, a wearable production device, or unrestricted autonomous repairs. The account separates the agent-written software from the owner’s physical and policy decisions.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Devpost · Aneesh Grover
Frequently asked questions
Did Claude do hardware bring-up?
No. The author says architecture, bring-up, and testing were done by hand.
How can approval be revoked?
The physical board supports a double-shake revocation gesture.
Were all demonstrated devices real servers?
No. The interface labels container stand-ins and the mock alert.
More details
About this account
Claude implementation sessions built the gate, executors, bridge, and dashboard under a shared interface contract. Grover brought up the FREE-WILi and Pixel hardware and corrected firmware and button-stream integration.
Opened original MHacks account. Claude Code wrote implementation; physical bring-up and tests were human. Gemini is the custom runtime agent, not the primary implementation tool. Sound, wearable form factor, and trust demotion have documented gaps. Suggested prompt written by Agents In Use; no original task prompt was published in the verified account.
Credit: Aneesh Grover — Devpost
Source published 4 Oct 2026
Source checked 14 Oct 2026
Tools and topics
Claude Code · FREE-WILi · Pixel 9 Pro · adb · SpacetimeDB · Gemini API
physical approval · rollback · device control
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.