SealGate lets Grok control a Flipper Zero from an Android phone

Why it matters

SealGate's demo shows MCP moving beyond software connectors. Once cloud agents can operate nearby hardware, permission controls and auditable tool calls become part of device security.

Eito Miyamura (@Eito_Miyamura) used SealGate to let Grok operate a Flipper Zero from an Android phone, demonstrating the setup on a hotel television in a video posted September 14th. The short demo turns a cloud chatbot into an interface for hardware sitting beside the phone.

https://x.com/Eito_Miyamura/status/2099453274915262888

Miyamura, who founded the London-based AI security startup SealGate, studied computer science at the University of Oxford and previously worked on self-driving infrastructure at Wayve. His startup, formerly called EdisonWatch, sells a control layer for monitoring and restricting how AI agents access company data and external tools.

The television demonstration puts a less corporate face on the same technical bet: agents will increasingly need to reach software and hardware that run locally, even when the model itself runs in a cloud service.

Miyamura described the result as getting Grok to "hack a hotel TV" and said users could speak or type commands rather than memorize the Flipper Zero's controls. The label is doing some promotional work. Flipper Zero already includes an infrared transmitter and a library of commands for televisions, air conditioners, projectors and stereo systems. Its official documentation describes the universal remote function as cycling through known infrared protocols until a device responds.

The new element is the agent in the control loop. Instead of selecting commands directly on the Flipper Zero, Miyamura exposed phone-level hardware controls to Grok through SealGate's implementation of the Model Context Protocol, or MCP. MCP gives models a standard interface for discovering and calling external tools.

From cloud model to local hardware

SealGate's open-source Android client runs a device-side daemon that maintains an outbound WebSocket connection to the SealGate gateway. Cloud agents can send MCP requests through that tunnel to modules running on the phone without an inbound port being opened on the device. The repository had 27 commits and 10 open pull requests when reviewed on September 14th.

The Android software exposes a restricted shell-like tool called Mobile Bash. Its documented commands cover device information, battery state, Wi-Fi, Bluetooth, USB and camera access. The Bluetooth module supports device scanning, pairing, GATT operations and serial connections, providing a route for an agent to communicate with nearby peripherals such as the Flipper Zero. Android's own permission prompts still apply.

Grok connects on the other side as a custom MCP client. SealGate's Grok setup guide instructs users to add the company's public gateway as a custom connector, after which Grok can discover the tools attached to the account. SealGate says custom connectors require a paid Grok tier.

That architecture extends work Miyamura and Ilia Manolov published on August 24th. Their stdio tunnel design was built to connect cloud agents with MCP servers running on personal computers, where local credentials and applications remain. The Android client carries the idea onto phones and adds direct access to mobile hardware interfaces.

The security startup widening the action surface

SealGate's sales pitch centers on controlling agents before they leak data or take prohibited actions. Its documentation says the gateway monitors tool calls, assigns risk scores and applies deterministic policies at runtime. Administrators can also quarantine newly discovered MCP servers and retain audit trails of agent activity.

Miyamura's Flipper demo shows why those controls become consequential once agents move beyond calendars, databases and messaging apps. A mistaken tool call connected to physical hardware can have an immediate external effect. An infrared command that changes a television channel is low stakes; the same interface pattern can reach Bluetooth devices, USB peripherals and other equipment exposed through custom modules.

SealGate is therefore making two bets at once. Miyamura expects cloud agents to operate local devices through natural-language requests, and he expects companies to pay for a gateway that records and restricts those requests. The hotel television is a compact demonstration of both propositions: the agent gains hands, while SealGate stays in the path between the prompt and the hardware.