Dogfooding a Voice Agent to Build Its Own LED Status Feature

I built a voice transcription tool for Windows. Then I used it to talk to an AI agent that built a new feature for the tool itself.

The Project: Windows AHK Dictate

Press a button, speak, get text. Runs entirely on your machine using Whisper. No API keys, no subscriptions, no data leaving your computer.

Windows AHK Dictate works with any mouse or keyboard that lets you bind buttons to keystrokes - Logitech G HUB, Razer Synapse, whatever. Press the button, speak, press again, and the transcribed text appears wherever your cursor is. Double-click to end and immediately submit - useful for agent chat.

The Ask: LED State Feedback

Once the basic tool was working, I wanted visual feedback. The G502 has RGB LEDs, and I wanted them to show recording state - idle, recording, processing.

What the Agent Did

I described what I wanted and let the agent figure it out:

1. Dialed into the Logitech LED SDK - Found the right DLLs and figured out how to interface with them
2. Queried hardware identifiers - Found my specific G502 among all connected Logitech hardware
3. Traced through C headers - The LED SDK isn’t well-documented. The agent worked out the function signatures from header files
4. One-shot programmatic LED control - Direct control over the LED state, not just preset profiles
5. Wired it to Whisper - Connected LED state changes to the transcription service lifecycle

High-level request to low-level hardware integration. One session.

RDP Support

I wanted this to work over RDP so I could transcribe on my home server from anywhere. By leveraging the Interception driver package, input gets captured at a low enough level that it works seamlessly over RDP.

I could have evaluated existing tools that might half-meet my needs. Instead, I just built one.

Try It

The project is open source: Windows AHK Dictate on GitHub

Works with other mice and keyboards too - the G502 LED integration is a bonus. Private voice transcription that works over RDP.

---

Building something cool with local AI tools? Drop a comment.