Matrix homepage recommendation
Matrix is a minority writing community. We advocate sharing real product experiences, practical experiences and thoughts. We will occasionally select Matrix’s highest-quality articles to showcase the most authentic experiences and opinions from users.
The article represents the author’s personal views, and the minority party only slightly modified the title and formatting.
0. Why do you want to equip AI programming with a handheld remote control?
Recently, more and more people use terminals to run AI programming (such as OpenAI Codex, Claude Code, WorkBuddy or various CLI Agents). After using it for a while, I can clearly feel that the biggest changes in the workflow are:Less typing, more conversations and confirmations。
In the past, most of the time when writing code was typing on the keyboard, but now it is often:
- Organize a piece of thinking and describe the requirements to the Agent in natural language;
- Watch it output code and Diff;
- Press Enter to confirm execution of the command, or press to interrupt midway;
- Switch between several different terminal windows or workspaces to see if each Agent has finished running.
At this pace, voice input is indeed a real improvement in efficiency – speaking is often much faster than typing a long prompt. But if you rely solely on the keys on the keyboard to wake up the voice, you still have to sit directly in front of the table and put your hands on the keyboard.
OpenAI’s hardware prototype some time ago Codex Micro After it was made public, many various “Vibe Coding physical buttons/small hardware” appeared in the community. But in reality, many DIY small peripherals are either rough in workmanship and have loose buttons, or they are early experimental toys with uneven stability and connection experience.
In fact, there is no need to buy those premium peripherals or bother with 3D printing. One that you can buy everydayXiaomi TV Bluetooth Voice Remote Control 2 Pro (Model RC003)the cost is about 50~60 yuan, it is mass-produced by a large factory, the mold is mature, and it comes with Bluetooth BLE and voice microphone.
Using it as a handheld controller, the core appeal is actually very simple:
- Push-to-Talk: Press and hold the voice button on the remote control with one hand to directly dictate the prompt, and let go to automatically display the screen;
- Blind cutting without keyboard and mouse: When leaning on the chair, use the direction keys with one hand to switch between the terminal Tab and Workspace, press the middle OK key to confirm release, and press the return key to interrupt.
This solution is not only suitable for what I use daily macOSthe community currently also provides Windows running path. The specific configuration methods and pit routing details are recorded below.
1. Hardware details: Why choose RC003?
Xiaomi’s remote control usually sells for about 50 to 60 yuan on second-hand platforms or e-commerce platforms. Choosing it mainly depends on several hardware details:
- Native Bluetooth BLE direct connection: It adopts the standard Bluetooth protocol and does not require an external USB receiver. The computer’s built-in Bluetooth can be directly paired and connected;
- Metal texture that far exceeds the price of tens of yuan: The biggest pain point of many low-priced small hardware or DIY buttons is that the shell is light, plasticky, and even has obvious burrs from 3D printing. As the original remote control for Xiaomi high-end TVs, RC003 has very solid workmanship:
- One-piece metal body: The champagne gold surface has been finely frosted, feels warm and cool and does not easily leave fingerprints. It has a certain sense of weight and balance in the hand, and is by no means a thin and flimsy plastic;
- Crisp and clear key sequence feeling: The inside of the button does not feel like a soft conductive rubber. When triggered, there is a crisp confirmation feeling (Clicky) similar to micro movement. The stroke is moderate and the rebound is sharp. It is clear whether the button is pressed or not during blind operation;
- Unobtrusive table look: Minimalist black and gold contrasting design, usually placed on the keyboard palm rest, solid wood desktop or digital table mat, it looks like a specially made handheld digital accessory, without any sense of disobedience;
- Button layout suitable for blind operation：
- The OK button in the middle has an obviousconcave arcyou can instantly locate it by placing your thumb on it;
- The outer ring direction ring is available in four directions: up, down, left, and right.Raised microscaleyou can clearly confirm the key position without looking;
- The microphone button on the top is independently centered. When holding it with one hand, the index finger rests on the bottom and the thumb naturally falls on the voice button;
- Near field vocal pickup: The built-in microphone of the TV remote control itself has been optimized for picking up human voices at close range. The environmental noise is relatively small, which is very suitable for daily oral prompts.
2. Software configuration: connect the system and voice input
Although the direction keys of the remote control can be used as ordinary input devices after Bluetooth connection, the voice keys on the remote control transmit dedicated voice data packets, and the operating system will not recognize it as a regular microphone by default.
Community open source HD838A/remote-mic-app（SayAll） Two things were solved:
- Generate a virtual audio input device in the system (under macOS
MiRemoteV 2ch）； - Monitor the voice events of the remote control. When the voice button is pressed and held, the audio is received and streamed to the virtual microphone in real time. It ends when the voice button is released.
In conjunction with voice input methods that support “Press and Talk (PTT)” (such asWeChat input method), you can press and hold to speak and release the screen.
2.1 macOS platform configuration process
- Pairing the remote: In the Bluetooth list of Mac system settings, press and hold the “Home button (house icon)” and “Menu button (three bars)” on the remote control at the same time, wait for the indicator light to flash, and click Pairing Connection on Mac;
- Install SayAll and audio driver：
- Download and install from GitHub Releases
SayAll.app； - After opening, follow the prompts to install the audio driver plug-in (located in
/Library/Audio/Plug-Ins/HAL/MiRemoteV2ch.driver）； - Open “System Settings ➔ Sound ➔ Input”, and you can see that the input device has appeared
MiRemoteV 2ch；
- Download and install from GitHub Releases
- Set up WeChat input method (WeType)：
- The speech recognition rate and punctuation segmentation of WeChat input method are very accurate, and it supports long-pressing shortcut keys to speak;
- Turn on “Hold the shortcut key to speak” in the WeChat input method settings and set the shortcut key to
Fn(or specify a key combination); - Sound input source selection
MiRemoteV 2ch。
2.2 Windows platform configuration process
If you are using it in a Windows environment (for example, using Windows to run a terminal, run Codex or WorkBuddy):
- Also press and hold the paired remote control in “Bluetooth and other devices” in Windows settings;
- Run the Windows preview version of SayAll (or the community’s Windows adaptation solution for RC003, such as
zhaozhuque/MIC-RC003-Windows）； - Windows Preview builds by default map voice shortcuts to
Ctrl + Win； - In the Windows WeChat input method, the voice input shortcut keys are also matched as
Ctrl + Winthe microphone can select the audio device mapped by the remote control.
3. Quick troubleshooting of several key details
There are several experience details that are easy to get stuck in during the actual configuration process. Dealing with them in advance can save a lot of time:
Tip 1: How to skip the SayAll boot interface if it gets stuck? (macOS)
When installing SayAll, for some macOS versions, after connecting the remote control, the novice guidance interface may remain stuck at the last step and cannot be clicked to complete.
No need to reinstall, just execute a line of command in the terminal to write the completion mark:
defaults write com.hd838a.RemoteMic "onboarding.completedVersion" -int 1
defaults write com.hd838a.RemoteMic "onboarding.step" -string "complete"Restart SayAll to directly enter the complete key mapping setting interface.
Tip 2: The Mac version of WeChat input method cannot input the independent Fn key?
On macOS, if you want to use a single press of the independent Fn key as a voice trigger, the WeChat input method setting interface occasionally fails to input normally. It is recommended to directly set an unpopular key combination (such as Control + Option + V) in the “Shortcut Key” of the WeChat input method, and map it to the same key combination in SayAll, with zero threshold and no conflict.
If you insist on using single-key Fn, the WeChat input method configuration is stored in MMKV with CRC32 verification. It can take effect by recalculating the verification through the following simplified script:
# 微信输入法 MMKV 配置修复 (重算 CRC32 校验)
import os, zlib, struct
path = os.path.expanduser("~/Library/Application Support/WeType/mmkv/wetype.settings")
data = bytearray(open(path, "rb").read())
open(path + ".crc", "wb").write(struct.pack("Then select the voice mode in SayAll as fn It works seamlessly.
Tip 3: No response when pressing the arrow keys quickly? Turn on “Allow continuous rapid presses”
After many users map the cross keys to custom shortcut keys in SayAll, they will find that “you can switch when you press it once, but nothing happens when you press it twice quickly and the second time.”
reason：
- SayAll enables anti-combination protection for custom shortcut keys by default, and will filter continuous input before the keys are stable;
- If you set “double-click” in the interface, the system will wait for tens of milliseconds to determine whether it is a double-click when single-clicking, causing significant delay.
Feel optimization settings：
- In the SayAll direction key settings,Clear all “double-click” bindings；
- Check the box below the corresponding button “Allow continuous rapid presses”（Allow Rapid Repeated Presses）。
After this setting, the key eliminates the double-click waiting window, clicks with zero delay, and responds instantly every time when clicked quickly in succession.
4. Terminal and Agent workflow configuration: Herdr / Codex / WorkBuddy
After getting the remote control, the most important value is to use it with the terminal multiplexer in the development workflow.
Here are the ones I commonly use Herdr(Terminal manager specially designed for multi-Agent collaboration, running in Ghostty or any terminal) as an example, compatible with OpenAI CodexTencent WorkBuddy (CodeBuddy) and common Agent tools such as Claude Code.
4.1 Key mapping for spatial intuition
In a multi-Agent scenario, the layout of the interface usually has a natural spatial logic:
- The top is arranged horizontally Tabs(Tabs usually hold different tasks or views in the same project);
- The left side is arranged vertically Workspaces(Workspace, such as running Codex, WorkBuddy and a test environment respectively).
We map the remote control to intuitive prefix-free shortcut keys:
| Remote control buttons | Send shortcut keys | Corresponding action | Practical use |
|---|---|---|---|
| Direction left ◀ | Ctrl + Alt + Left |
PreviousTab | Move your thumb to the left to return to the previous code or output window |
| Direction right ▶ | Ctrl + Alt + Right |
NextTab | Move your thumb to the right to view the parallel terminal output on the right |
| direction▲ | Ctrl + Alt + Up |
Previous Workspace | Switch up the workspace (such as switching to the project running Codex) |
| Direction down▼ | Ctrl + Alt + Down |
NextWorkspace | Switch down the workspace (such as switching to the execution area of WorkBuddy) |
| Center OK button | Enter |
Enter to confirm | When the Agent proposes operation confirmation, click to release it directly. |
| Return key⮌ | Ctrl + C / Esc |
Interrupt/Cancel | If you find that the Agent is running deviating from expectations, click to interrupt the current round. |
| Voice key 🎙 | Fn (Long press PTT) |
Voice input | Press and hold the dictation prompt, let go and the text will automatically fall into the command line |
4.2 Keep the original keyboard habits: Herdr’s multi-key array
If you are still used to tmux style on the keyboard Ctrl+B then press p / n Cut tags in Herdr’s configuration file (~/.config/herdr/config.toml) supports using arrays to declare two sets of keys at the same time:
[keys]
previous_tab = ["prefix+p", "ctrl+alt+left"] # 左/右键：切 Tab
next_tab = ["prefix+n", "ctrl+alt+right"]
previous_workspace = ["prefix+up", "ctrl+alt+up"] # 上/下键：切工作区
next_workspace = ["prefix+down", "ctrl+alt+down"]Execute after modification herdr server reload-config Hot loading will take effect. In this way, the keyboard shortcuts can still be used when typing while sitting at the desk, and when leaning on the back of the chair and holding the remote control, you can cut straight with one hand without prefixes.
4.3 Actual cooperation in Codex/WorkBuddy
When you run different Agents in multiple workspaces in parallel:
- In Codex: It is often necessary to confirm the generated changes multiple times. Press and hold the voice key with your thumb: “Don’t refactor this part, just add a null value check to the original function.” Let go and the text will appear on the screen. Click OK in the center to submit; if Codex asks whether you are allowed to execute the bash script, click OK again to release;
- In WorkBuddy / CodeBuddy: When multiple Agents are performing separate tasks, the left hand can completely leave the keyboard, and the right hand holds down the remote control to issue the task: “Help me generate the API document just now into Markdown format.” Then use the up and down keys to switch to other workspaces to check the compilation log.
5. Daily use experience
After using this solution for a period of time, the most real change is not an earth-shaking change in the way of working, butFeeling more comfortable physically：
- Long Prompt obviously makes me more willing to talk.: In the past, when typing three or four sentences of requirements or boundary conditions, I was often too lazy to type and condensed it into half a sentence, causing the Agent to understand inaccurately. Now, hold down the remote control to naturally speak the complete context of one to two hundred words, the recognition accuracy is very high, and the communication cost has been reduced;
- Multi-terminal inspection is more convenient: When several agents are running tasks, holding the remote control in your hand and pressing the direction keys twice to check the status is much easier than moving your hand to the touchpad or mouse to switch screens;
- Hardware cost is almost no psychological burden: Compared with specialized small hardware that costs hundreds of dollars, Xiaomi remote control that costs tens of dollars doesn’t hurt if it breaks, and the mold workmanship is completely online.
If you regularly use various AI programming tools, you happen to have an idle Xiaomi Bluetooth remote control on hand, or you can buy one for tens of dollars and spend more than ten minutes to configure it. It will be a low-cost but solid experience for a small peripheral upgrade.
> Follow the minority public account to unlock a new reading experience📰
> Practical and easy-to-use genuine software, presented to you by the minority 🚀
