21:19Sharbel A. · Talking Head
100 Days With Hermes Agent in 21 Minutes
Five mistakes that keep you using an AI agent like a smarter search box — and the systems that fix each one.
June 17thThe argument in one line.
Claude Code can already reach your codebase and every connected app through MCPs, but it can't see your screen or be summoned from another window, and a $30-a-month tool called VoiceOS closes both gaps with one global hotkey.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude Code daily and spend real time translating visual bugs into written descriptions.
- You want a global hotkey that reaches Claude Code from any window without opening a terminal.
- You're weighing a paid monthly tool against the free voice options already built into Claude.
SKIP IF…
- Claude Code isn't your primary coding agent, since the entire pitch is built around it specifically.
- You've built your workflow around local inference and won't consider a cloud-only tool.
The full version, fast.
Claude Code can read an entire codebase and connect to email, Slack, Notion, and calendars through MCPs, but it can't see what's on screen and can't be reached outside its own chat window. VoiceOS, the paid tool this video is sponsored by, closes both gaps: a global hotkey lets you circle anything on screen, describe the change by voice, and fire off a Claude Code run from any app without opening a terminal. It costs $30 a month, or $12 a month billed yearly, after a seven-day trial, then drops to limited free use. Its real weaknesses are that it's cloud-only and its connector list is thin outside major apps. Three free alternatives exist (ChatGPT's voice mode, Claude Code's built-in mic button, and the open-source Voice Mode MCP), but none of them can hand a circled screen element straight into a running Claude Code session.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:59
01 · Giving Claude Code Eyes
The recurring screenshot-crop-paste-describe workflow, and why it's slow and error-prone.
00:59 – 02:02
02 · What Claude Code Still Cannot Do
Claude Code already reaches email, Slack, Notion, and databases through MCPs; its real gap is no eyes and no doorbell.
02:02 – 03:50
03 · Why I Tested VoiceOS
How the sponsorship came about, the two-week testing window, and VoiceOS's pricing and trial terms.
03:50 – 04:44
04 · Complete Setup in Three Minutes
Install, create an account, then turn on the apps tab and connect Claude Code.
04:44 – 05:40
05 · Circle Anything on Your Screen
Holding the hotkey while circling an on-screen element turns it into visual prompt context.
05:40 – 07:01
06 · Stop Opening the Terminal
The hotkey fires a Claude Code run from any window; a side panel tracks every run in progress.
07:01 – 08:59
07 · Build Your Own Integration
Building a custom Hermes agent connector for a tool with no default integration.
08:59 – 10:02
08 · Where VoiceOS Breaks
Three limitations after two weeks: cloud-only, a thin connector list, and dictation mishears.
10:02 – 11:34
09 · The Best Free Alternatives
ChatGPT voice mode, Claude Code's built-in mic, and the open-source Voice Mode MCP, and what each one lacks.
11:34 – 13:05
10 · My Verdict After Two Weeks
Who should take the free trial, and who should stick with the free alternatives.
Lines worth screenshotting.
- Claude Code can already reach your email, Slack, Notion, and calendar through MCPs, so its real limitation is that it can't see your screen or be summoned from outside its own chat window.
- VoiceOS costs $30 a month billed monthly or $12 a month billed yearly, with a seven-day free trial before it drops to 100 dictation sessions and 25 agent sessions a week.
- Circling an object on screen while holding a hotkey turns that circled element into visual context for a prompt, replacing a written description like 'the third box, no, the text under that.'
- A held hotkey works from any window, so a bug spotted on a live website can be fixed by circling it and speaking, without switching to a terminal or code editor.
- Every voice-triggered run appears in a side panel showing which model and reasoning intensity it's using, and a run can be interrupted or redirected mid-task by holding the hotkey again.
- When no built-in connector exists for a tool, VoiceOS can generate a custom integration from a plain-language description, which is how the reviewer built a Hermes agent connector.
- VoiceOS is cloud-only with no local inference option, which rules it out for anyone who has built their workflow around running models locally.
- The connector list is strong on mainstream apps like Mail, Slack, Notion, and Drive, but thin everywhere else, so niche tools usually need a custom-built integration.
- Voice dictation can mishear technical or invented product names, and the fix is adding those words as keywords inside the app's dictation settings.
- ChatGPT's desktop voice mode can see your screen and talk while you work, but it can't let you circle a specific element and it doesn't help if Claude Code is your actual coding agent.
- Claude Code already has a built-in microphone button for dictation, but it only works while you're already inside the Claude Code window, with no screen-sharing at all.
- Voice Mode is a free, MIT-licensed MCP server that gives Claude Code local speech-to-text, so audio never leaves your machine, but it has no screen awareness and takes more setup work.
A Hotkey That Gives Claude Code Eyes
VOICE AND SCREEN CONTROL
A global hotkey plus screen-circling closes the two real gaps in Claude Code, seeing your screen and being reachable from any window, though the tool is worth trialing free before paying for it.
01Giving Claude Code Eyes
- The recurring workflow of cropping a screenshot, switching windows, pasting it in, and writing out which element to fix is slow and often points the agent at the wrong thing.
- Claude Code has no way to see what's currently on your screen, so anything visual has to be translated into words first.
02What Claude Code Still Cannot Do
- Claude Code's connectors (MCPs) already let it reach your inbox, calendar, Notion, and databases, so the real gap isn't reach, it's that it can't see your screen or be summoned from another window.
- You can talk to the Claude desktop app by voice, but Claude Code itself has no voice and no screen access built in.
03Why I Tested VoiceOS
- VoiceOS is a paid, closed-source tool: $30 a month billed monthly, or $12 a month billed yearly.
- A seven-day free trial is available, and canceling after it doesn't cut off access completely, it drops you to 100 dictation sessions and 25 agent sessions a week, about three agent runs a day.
04Complete Setup in Three Minutes
- Installation is download, open, and create an account, then two things need turning on before anything works: the apps tab (connectors) and Claude Code itself.
- Each connector uses its own OAuth approval window, and niche tools not on the default list may need a custom-built integration.
05Circle Anything on Your Screen
- Holding the hotkey while circling an element on screen turns that circle into visual context attached to the spoken prompt, replacing descriptions like 'the third box, no, the text under that.'
- The same hotkey works from any window, including a live website, not just inside a code editor.
06Stop Opening the Terminal
- A voice request opens its own Claude Code instance and starts working immediately, without opening a terminal or the desktop app first.
- A side panel shows every run, which model and reasoning intensity it's using, and lets a run be stopped or redirected mid-task by holding the hotkey again.
07Build Your Own Integration
- When no built-in connector exists for a tool, a custom one can be created by describing what's wanted in plain language and refining it through a short back-and-forth.
- Simple integrations, like a calculator, can be one-shot, but connecting to a tool running on a separate device took real trial and error.
08Where VoiceOS Breaks
- It's cloud-only with no local inference option, which matters if your workflow is built around running models locally.
- The connector list covers mainstream apps well but is thin outside them, so most niche tools mean building a custom integration.
- Voice dictation can mishear technical or invented words, fixable by adding those terms as keywords in the dictation settings.
09The Best Free Alternatives
- ChatGPT's desktop voice mode can see your screen and talk while you work, but you can't point at specifics, and it doesn't help if Claude Code is your actual coding agent.
- Claude Code's own microphone button gives voice input, but only from inside the Claude Code window itself, with no screen access.
- Voice Mode, a free MIT-licensed MCP server, gives Claude Code local speech-to-text so audio never leaves your machine, but it has no screen awareness and takes more setup.
10My Verdict After Two Weeks
- The reviewer reports reaching for the tool constantly, including for tasks outside Claude Code, like sending emails or setting Notion tasks while doing something else.
- The recommendation is conditional: try the free trial if you're in Claude Code daily and a meaningful share of your work is visual, not as a default purchase.
Terms worth knowing.
- MCP (Model Context Protocol)
- The connector standard that lets Claude Code reach outside tools like email, Slack, or a database, instead of only reading local files.
- Hotkey run
- A voice-triggered request sent to a coding agent by holding a keyboard shortcut, without first opening the agent's own chat window.
- Agent session
- One unit of usage in VoiceOS's pricing plan, counted each time a voice request triggers a Claude Code (or other agent) run.
- Custom integration
- A connector built inside VoiceOS for a tool that isn't on its default app list, created by describing the desired behavior in plain language.
Things they pointed at.
Lines you could clip.
00:20
“You've done that today and you'll do it again tomorrow.”
06:40
“It becomes something you throw work at while you carry on doing something else entirely.”
12:04
“At my scale with the business that I'm running, it's a drop in the ocean compared to what it lets me actually do.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
You're building something in clot code, but something on the page looks wrong. So you crop a screenshot, you save it, you switch windows, you paste that in, and then you write a paragraph explaining which button you actually want to fix. You send that message and it changes the wrong thing.
And it also breaks something else entirely. You've done that today and you'll do it again tomorrow. Clot code can read your entire code base.
It can reach your email, your Slack, your Notion, but it can't see the thing you're looking at right now. And it can't be reached from anywhere except a chat box. What if I told you that there's a way to solve both of those problems with one tool?
This is exactly what I'm going to show you in this video. We're going to give clot code eyes and ears across our entire workspace. I'll also cover what the tool is, where it's falls apart and how you can start using it for free.
So let's get started. Okay, so let's be precise about what cloth code is actually missing. Cloud code is not trapped in your folder.
If you use its connectors, aka its MCPs, it can reach your inbox, your calendar, your notion, and your database. So reach isn't the problem here. The gap is much smaller than that and a little more annoying.
It has no eyes. It cannot see what's on your screen right now. So anything that's visual, you have to translate into words.
And words are often a terrible form. format for that box, the third one, no, the text under that. And it's got no doorbell.
You can't reach for it whilst you're on another window, for example. You can summon regular Clod from anywhere, but not the world's most powerful coding agent. Instead, you have to open the Clod desktop app, go to Clod code, be in the right folder, and ask for the thing you want.
That problem right here is exactly what I've been looking to solve for months. I get a lot of emails from companies asking me to look at their product. And honestly, I end up saying no to almost all of them.
But I recently got an email for the tool I have literally been looking for for a while now. And here's why that is. The Claude app on your desktop has voice.
You can talk to it. But I barely use the Claude app. I mean, I live inside of Claude Code.
And Claude Code has no voice or no screen access. Nothing that remotely looks like that. So I've had the thing I wanted sitting one app away from the thing I actually work in for months.
So I agreed to take a look at the product and I have been testing it out for the last two weeks now. So yeah, this video is sponsored by Voice OS. They gave me complete access to the tool for free and told me I could be as honest.
as I want about it, which is great because you're going to see me cover this tool's best features, but also its downsides. The catch is that this is a paid closed cloud tool. You can get started with the seven -day free trial and decide if you like it from there.
It is $30 a month if you pay monthly or $12 a month if you go yearly. And if you cancel after the seven -day trial, then you don't lose access completely to the tool. You drop down to 100 dictation sessions.
and 25 agent sessions a week. That's about three agent runs a day, which is enough to keep using it, just not enough to work the way I do in this video. I did include free alternatives at the end of this video in case that's what you want.
However, each have their own limitations, which I will also cover there as well. But with that out of the way, let's actually install it and test it out together. Installing it is simple.
You go to their site, you download it for you. your operating system. You open it and you make an account.
That is the whole install. Then there are two things to turn on before any of this works. The first is the apps tab.
That's where every connector lives and you turn them on one at a time. Each one opens an OAuth window. You approve that and it's connected.
You turn on clot code, then you click into it and you connect that. The connector list covers the most common apps like Mail, Slack, Notion, Drive, Canon. but if your workflow lives in something a little niche, or if the app you need is not already built into that list, you may need to add or build an MCP integration, which we will get into how to do that later in this video.
And with Clothcode connected, here's the feature I actually made this video for. Now, this is the part that changed how I work with Clothcode. You hold the hot key like this, you can see it activated.
And while you're talking, you can circle something on your screen that it will give us context to your prompt. And that circle that we drew, as you can see, just became context and just became the thing that it's pointing to visually inside of your prompt.
So instead of saying something like change the color from red to the one below that. Text where it says you can. What you can literally do is say, hey, Claude, could you change this red text to yellow for me?
Changing the red text to yellow. And that's it. Just like that.
It sent it its request. It opened a Claude instance, a Claude code instance, and it's currently doing it. This is a big improvement from taking a screenshot, opening the desktop app or the terminal, pasting the screenshot inside and describing the specific thing you want to change.
Here's what I absolutely love about this feature. The hotkey works everywhere. You're on your website and you notice a bug.
For example, over here, it's not showing the thumbnail for the next articles coming up. I can literally just hold the hotkeys, circle around the problem and give it my prompt. Hey, Claude, could we add thumbnails next to each title here on my website?
We're working on that. There you go. It just sent its request and it doesn't just fire and forget.
There's a panel on the left where every. run lives. There's another panel at the top where you see your current run.
You can see which model it's on. I can click into it. I can see which model it's on, which intensity it has.
I can modify that. I can ask it to stop if I don't want it, or I don't know if it misunderstood me. I can look at the prompt and reread it.
And you can even hold the hotkey mid run to ask it to change something. Actually, nevermind. I wanted to design a specific thumbnail instead for each video.
Just like that. So clot code stops being a place you go and sit and it becomes something you throw work at while you carry on doing something else entirely. And once you've got that going, you can then go one level further and build integrations for any tool on your desktop.
Let me show you how. Another thing I love about this thing is that if you can't find an integration for a tool that you use, you can just create one. For example, I love Hermes.
agent, but there isn't an integration for it there. So I went to custom integrations and I built one myself. I clicked on create and then I just explained what it is I wanted to do.
Hey, I want to create a Hermes integration. And then I had a little back and forth with it until this was done right here. This means you can build integrations for this tool to be able to reach any tool with a usable API or MCP path.
so that you can use that tool with VoiceOS from anywhere on your desktop. Now I can just run Hermes and ask it something like, Hermes, what tasks did you complete today? One sec.
Checking task. I have sent the question to Hermes agent and the summary will appear in your panel as soon as it finishes. There you go.
And it will just answer me from anywhere. I don't need to go to Telegram or the Hermes desktop app anymore when I'm busy doing something else. Hermes completed email audits and partner replies, updated Notion metrics, rescored X posts, checked YouTube competitors, and verified all 10 jobs, while demo assets remain outstanding.
There you go. It's that simple. Building custom integrations for more complicated apps isn't one shot though.
For example, I built a calculator integration and that was literally as simple as asking it to do that. But creating a custom Hermes integration took a bit of back and forth given my Hermes instance runs on a separate device and I want to be able to control it from an entirely different device instead. By the way, I will include the custom Hermes integration I built.
for free in the description, along with everything else I cover in this video. And with that, let's talk about the places where this tool can fall a little short. Every tool has a breaking point.
And after I tested this tool for the past two weeks, here are mine. The first and the biggest one in my opinion is that it's cloud only. There's no local inference option.
If you've built your whole setup around running things locally, this might bother you. as well the second is the connector list it's good on the obvious apps but it's thin outside of these.
The good thing is I literally only needed one more integration, which is Hermes agent, and then I was able to just build a custom one for myself in under an hour. The third is that it can sometimes mishear technical words, like I would say Hermes and it would hear Hermes, but luckily there's an easy fix for that. You just go to the app, go to dictation, and add all the keywords you typically use so that it familiarizes itself with them and hears those more accurately.
And this brings us to the question I'd be asking if I were watching this video. What are the free alternatives? There are free ways to get some of this and I thought I'd do some research and dig into each one.
I do have to warn you, though, every one of these alternatives has their own limitations, which I'll be covering as well. The first one is ChatGPT's desktop app, which can also use voice mode for codecs. If you already use ChatGPT or codecs for your coding, you can enable the voice chat hotkey.
It can even see your screen while you're talking to it. The only downside of that is you can't point at specifics on your screen using that, and it doesn't really solve the problem if clot code. is the coding agent that you use.
The second one is already sitting inside Cloth Code. You just click on the microphone button and say what you want to say. But you have to already be inside of Cloth Code.
You can't activate it when you're on another window and you can't show it your screen at all unless you manually screenshot and paste that into the chat window. The third is Voice Mode. It's an MCP server that gives Cloth Code voice conversations.
It is MIT licensed. and it can run speech to text locally. So your audio never leaves your machine, but it is voice only.
There's no screen watching or screen sharing and getting it set up takes a little bit more work. So here is the honest comparison. The free options are strong on voice.
What none of them can do, though, is hand that specific piece straight from your screen onto a cloud code run that's already working inside your project. So after using this tool for the last two weeks.
Here's what I think about it. I have noticed myself reaching out for it a lot more frequently. I mean, it's just so so convenient and I understand the price is not going to be convenient for everyone watching this video but at my scale with the business that I'm running it's a drop in the ocean compared to what it lets me actually do and how convenient it makes working for me.
I've even started using it beyond Clothcode and Hermes. I've noticed myself reach for it and ask it to send me emails whilst I'm editing a video or set me a task on Notion whilst I'm on a meeting. I'm even using the dictation mode whilst I'm on random tabs because it almost feels like a real life Jarvis to me.
So if you're in cloud code every day and a lot of what you do is visual then take the free trial and try it out yourself. Tell me in the comments by the way what's the one thing you wish your setup could do that it currently can't. Maybe I'll turn that into a video next.
Also as I mentioned the custom Hermes integration and everything else that I covered in this video are free in the description below. And if you enjoyed this video then leave a like or if you're new to my channel then subscribe because I have got a ton more content like this coming your way.
Oh and would you look at that the algorithm gods have decided you're very likely to enjoy this next video as well. So click it and maybe I'll see you then.
The Hook
The bait, then the rug-pull.
You crop a screenshot, switch windows, paste it in, and type out which button you actually meant, and half the time Claude Code still changes the wrong thing. This breakdown follows a two-week test of VoiceOS, a paid tool that lets you circle whatever's on screen and talk your fix straight into a running Claude Code session.
Named ideas worth stealing.
09:18list
VoiceOS's Three Breaking Points
- Cloud-only, no local inference option
- Connector list is thin outside major apps like Mail, Slack, and Notion
- Voice dictation can mishear technical or invented product names
The reviewer's honest limitations list after two weeks of daily use.
Steal forEvaluating any cloud-based AI tool before committing to a paid plan.
10:08comparison
Three Free Alternatives to VoiceOS
- ChatGPT desktop voice mode: sees your screen and talks, but can't point at specifics and doesn't help if Claude Code is your coding agent
- Claude Code's built-in microphone button: voice input, but only works once you're already inside the Claude Code window
- Voice Mode (MIT-licensed MCP server): free local speech-to-text, but no screen awareness and more setup work
Free ways to get partial voice or screen access to Claude Code, each with a real gap.
Steal forTesting the workflow before paying for a dedicated tool.
How they asked for the click.
FROM THE DESCRIPTION
AFFILIATECommission earned if you click.
OTHER LINKSAlso linked in the description.
