Engineers are coding out loud

Estimated reading time: 7 minutes

Key takeaways:

• Voice coding has gone mainstream. Claude Code, Codex, and Cursor all launched voice modes this year.
• Voice suits long prompts as speaking gives agents more context faster, though models can miss nuance.
• Self-consciousness is the biggest barrier: thinking out loud feels exposed.

---

Engineers and vibe coders everywhere are building software by ‘voice coding’ – talking out loud to coding agents, no typing required.

Callum Reid initially felt “kind of goofy” with a big microphone set up on his desk, hunched over and whispering into it all day. Then the other engineers on his team started asking about it. Soon enough, most of them got their own microphones, too. 

Reid, founding engineer at startup Coval, first started coding by voice around three years ago while dealing with a repetitive strain injury (RSI), using a model he fine-tuned to recognize his own whispering voice. Today, thanks to the growing availability of voice capabilities designed for engineering workflows, he’s far from alone.

Many voice coders aren’t even sitting at a desk, but rather coding via earbuds while commuting or walking their dog (including Reid, who says his dog has learned that when they’re walking, it’s not her he’s talking to).

Getting comfortable talking out loud to the computer can be a hurdle, but overall, those who are embracing voice coding say it enables them to work faster, work on-the-go, and is particularly useful for info dumping on AI agents.     

“For a long time I’ve been trying to do this, and in the last year, it’s really taken off as a main way I can engage with the computer,” he says. “I find I rarely actually have to go back to the keyboard at this point.” 

The voice coding tech stack

Claude Code, Codex, and Cursor all rolled out a voice mode in the last year. Like Reid, many engineers were already – and are continuing – to rig up their own setups

Alon Sabi, a developer and director of IT at a school who now regularly works via voice, programmed a Flick smart home device to function as a simple voice control button for his laptop. The button sits on his desk and allows him to simply tap it once to prompt the system to start listening. A double tap makes it return and enter, and after he’s given his prompt, another tap stops the listening, he says. Sabi also regularly uses Claude Code’s Remote Control feature to work via voice from his phone while out with friends or doing other things. 

Working only via voice, he recently built a digital PA system for the school using a Raspberry Pi as well as a replacement for the old audio and visual system the school uses for assemblies. While he estimates each project would’ve taken weeks coding in the traditional way, he says they took only hours coding with agents via voice. 

“It’s mind blowing how much you can do without even having a computer in front of you,” he says.

One of the main reasons Kirill Dubovitskiy built Happy Engineering, an open-source mobile app for orchestrating various agentic coding tools, was to add a voice mode before it was a feature in mainstream AI-coding tools.

“Basically, I wanted to be able to code from your phone,” he says. “And it’s a small screen – you cannot read as much code. So I think most of what you can really do is just provide more input to the agent.”

Of Happy Engineering’s 1,000 or so daily active users, around 70 start a voice session each day, according to Dubovitskiy. He knows some are hardcore users of the feature and talk out loud to agents all day, he says. It’s currently a paid feature, though they’re considering making it free. 

Beyond his own tool, Dubovitskiy has also experimented with different voice software like Talon Voice and Aqua Voice, particularly while dealing with an RSI of his own. 

Developers dealing with injury and disability have actually been creating ways to code by talking out loud to their computers long before large language models (LLMs) and AI agents. 

For example, in 2013, developer Tavis Rudd gave a famous presentation about how he built a system to code using customized voice commands, which he continued using even after recovering from his injury as he said it enabled him to code more efficiently than he ever had by hand.

Years later, a game developer made his own version while dealing with an RSI, telling Ars Technica he hoped his success with voice coding would promote it as an alternate way of working for others with similar diagnoses or disabilities.

Pros and cons of voice coding

The more recent boom in interest around voice coding isn’t stemming solely from necessity, but rather from a desire for speed. Plus, programming has changed. Whether you’re talking or typing, giving directions to agents in natural language has made it undeniably conversational. 

“What’s really nice about voice is you can flesh out the details and give extra context really effectively in a way that humans are super used to doing,” says Reid.

Richard Menning, a developer in Germany, started using voice to code a few months ago, first with Cursor and more recently with Claude Code. He also recently built a side project from his phone while walking his dog. Overall, he estimates he uses voice mode around 10% to 20% of the time, specifically when he wants to relay a lot of information and already has a clear idea of what he wants to say. 

“I can just speak faster and give more instructions or more detailed instructions than if I type it,” he says. 

Similarly, while Dubovitskiy says he uses voice at least 50% of the time while doing programming work, he uses it for basically 100% of long inputs because he finds that’s where it really delivers the benefits of speed.

“For very long inputs, it’s almost guaranteed I’ll use voice mode. Especially if it’s something I don’t need to think through, where the answer is obvious to me,” he says.

While effective for getting all the information out of his head and to the agents faster, Dubovitskiy finds these systems still struggle with prioritization and fully grasping the nuances of longer inputs. This challenge isn’t necessarily limited to voice mode, but it may exacerbate it, since voice mode seems to be particularly useful for and may even encourage the use of longer, more complex prompts.  

Reid finds that background noise can sometimes pose a challenge. Additionally, Menning has noticed the models don’t always correctly capture what he says, potentially because he’s not a native English speaker. Yet, both say they’ve seen the voice capabilities drastically improve and become more reliable in just a short period of time. 

“I will still use a screen to view results, and you still want to validate really sensitive stuff with the manual click of a button,” says Reid. “However, from beginning to end, through things like ‘go ahead and make that fix, merge it, ship it, and tell me when it’s done. Then go test it in production.’ It’s all done through conversation.”

A weird, naked feeling

The biggest challenge for developers exploring voice coding is the transition from thinking in their heads to thinking (and talking) out loud. 

Dubovitskiy notes that when you’re writing, you essentially have a canvas and get immediate visual feedback on how your thoughts appear. Writing is where a lot of thinking happens, but that process ceases to exist with voice coding, requiring a mindset shift from feeling the need to craft the right words to being willing to just let everything go.

“You kind of get time as you’re typing to think about what’s next, and I feel when you’re speaking, because the pace has increased a lot, you need to just brain dump. It’s stream-of-consciousness, and it’s a different type of input,” he says. 

All the engineers we spoke to also cited social etiquette and just feeling weird about talking out loud as a major hurdle. This isn’t limited to developers, either: the addition of voice capabilities to general chatbots and rise of tools like WhisperFlow is inspiring a whole new category of tech solutions designed to make voice interactions more discreet  – from earbuds that can pick up the slightest of whispers to sound-dampening masks.

Menning and Sabi both work remotely and expressed feeling awkward voice coding even in their own homes, especially when their partners are around. 

“At the beginning it felt wrong. It felt weird because my wife will sit in the living room, and we have a small house, so she hears me talking to the computer,” says Sabi.

Similarly, Dubovitskiy’s main interest in voice coding came from wanting to work while on-the-go, but he found that in reality, when coding from his phone, he was usually in public places. Ultimately, he just did not feel like the boost he realizes from getting his words out faster through voice was worth talking to his phone in public. 

Overall, he believes that self-consciousness is likely one of the biggest reasons more people aren’t voice coding, at least for now. Especially in an office, working out loud makes it possible for everyone to judge your process rather than your final output, like letting everyone read your first draft.  

“It’s just like this weird naked feeling of, oh f*ck, I’m just literally telling everyone what I’m thinking,” he says.

Get articles like this in your inbox

Choose your LeadDev newsletters to subscribe to.