# I Built an AI Chief of Staff with Hermes Agent (Full Setup)

Channel: Rakhul Karthick
Video: https://www.youtube.com/watch?v=pGIzhiZww9Q
Duration: 27 min
Language: English
Words: 4685
Transcript page: https://viewrankai.com/tools/youtube-transcript/pGIzhiZww9Q

---

[0:00] Take a look at this. So, I just sent one message in my Telegram and that kicked off an entire team. There's one moving into progress and that one is already done because I'm not talking to one agent here. I've got a chief of staff set up and he has got an entire team under him. One of them is doing the research, one is writing, one is checking the work before it comes back to me and I didn't tell any of them what to do. I just talked to my chief of

[0:24] staff and he worked out who does what. Take a look at this. This turned up at 8:00 a.m. too. My calendar, my urgent emails, what my team said overnight. I was asleep. This is Hermes Agent. So, it runs on my own computer. It remembers how I like to work and it gets better longer I use it. Today I'm taking you through from downloading it to this. One chief of staff, an entire team under

[0:49] him. Everything running on schedule. I'll take you through every step, every prompt, nothing is skipped. Hi there, I'm Rakul. If you're new to this channel, hit that subscribe button. Now, let's get straight into it. A quick bit of context before we start. A model like ChatGPT or Claude Opus is an engine. It's powerful, but on its own it just sits there. No wheels, no steering, no way to actually go anywhere. But Hermes is like a rest of the car. It's what wraps around the model and let it do

[1:18] things for you. It runs on your machine. It can use your apps and you can message it like you would message a real person. Now, here's the part that makes it different from something like Claude code. Every time it finishes a job, it looks at what worked and what didn't and it saves it as a skill. And a skill is just a markdown file sitting in a folder on your computer. So, the next time you ask for the same kind of thing, it doesn't start from scratch. It reads the file it wrote last time and that's why people describe it as an employee rather than a tool. It's not a one fancy

[1:53] feature. It just gets more useful the longer you leave it running. Two more things worth knowing. Number one, it can run completely local. For example, on your dedicated Mac mini or on your computer. You can also host it on cloud using the sites like emergent or Hostinger. And number two, it's model agnostic. So, you're not locked on to one company only. And that's what makes the chief of staff thing work. Because it remembers how I work, I can give it to one instruction instead of 10 and it fills the rest itself. That's the difference between managing an agent and managing a real person. And now, here's

[2:29] my honest bit before we start. You're still in charge here. It's not a robot you hand your life to and walk away. You give it access slowly and you check its work and you don't let it near anything that costs money or sends messages until you trust it. I'll show you safety switches as we go. Now, let's get

[2:47] started. Alrighty, so now let's install it. Let's go to Hermes agent.nowadaysresearch.com. So, this is their official site and you have got an option to download it to your Mac or to your Windows PC. So, here you can just click on desktop app and alternatively, you can also use your terminal to install it and you have just got the command here. Just copy, open your terminal and then install it. But for now, I will be using the Mac desktop version and just I will just save. And then, I will just click on Hermes drag and drop the applications and boom. So, this is the desktop app for the Hermes

[3:33] agent. And the first thing that you need to do is to give Hermes the brain. So, basically you would need to connect an LLM model. It can be your chat GPT, your Claude or Grok or whatever. So, you just need to give itself a brain. So, you have two options here. The first one is basically you can subscribe to the Hermes Agent subscription. So, they have

[3:57] got like $20 plan or $100 or $200 plans. If you just get the $20 plan, you would get access to all the top models. Uh there are like 200 plus models that you would have access to. But what I have done is I've got my ChatGPT subscription. I've got a plus plan, so I just use that as the brain for my Hermes Agent. You can also use Claude, but if you use Claude, you need to purchase additional credits on top of your existing plan. So, I just use ChatGPT, and it's pretty straightforward. To use your ChatGPT subscription, you just need to go into terminal and then just paste

[4:39] this command Hermes model and then click on enter, and you have got all the options that you have, and you just need to select the model that you want to choose, and just click on enter, and then you have two options here. One is to use OpenAI Codex, and then the other one is to use OpenAI API. I would choose OpenAI Codex, click on enter, and I've got some options here. One is to use your existing credentials, and the other one is to reauthenticate, and the other one is to cancel. But if you're setting up for the first time, you would have login. You would need to select login,

[5:15] and then it will take you to the authentication page in your browser. So, need to do is just to sign in using your ChatGPT account, and boom. And so, once you're done, you'll be back here to this terminal, and you would need to pick the model. So, it has got all the ChatGPT models in here. I'll just click on GPT

[5:35] 5.6 Soul, click on enter, and that's it. So, it says default model has been set to GPT 5.6 Soul via OpenAI Codex. That's all. So, now just click on new session, and now you have your models here. So, I will just test it. I will just tap on hi, and that's it. So, it's working fine. It just responded like, how can I help you today? Uh if you wanted to connect to Grok, I just do the same thing again, go back to terminal, and then just type in Hermes model, click on enter, and then I just select XAI Grok, and uh for this you need to actually

[6:13] purchase API credits or you can use Super Grok or premium plus plan, and then I'll just use this one. So, it just opened the browser, and then uh it just pasted the code that was displayed here, and click on authenticate, and uh and you would get something like this as device authorized, and that's it. And now you have to choose your model. So, you have lot of options here. So, I would choose the Grok 4.5, and that's it. So, the

[6:45] default model is now set to uh Grok 4.5. Now, I can just choose the Grok model if I want to. I just click on here, and um select Grok. Okay, so this is connected, too. So, I got a response on how it can help me today, and boom. So, we have connected uh ChatGPT and Grok to the Hermes agent. Now, I would show you how you can choose your follower model. So, just go on to settings, and then click on model, and you have the option here, and if you scroll down, you will have a fallback model option. You click on this, and then I will just click on

[7:24] Grok, and uh click on Grok 4.5, and that's it. My preference is to use GPT as my first model, and use Grok as the fallback model. Uh one important thing that I would like to highlight here is to store all your API keys in a .env file, uh which is inside the Hermes folder. Never put that directly in the chat because anything you type in the chat gets sent off to the model and it can be saved into your session history as a plain text. Just in case if your chat gets leaked or if you get to share your screen and then you would expose your API keys up to the public and

[8:04] that's very risky. So it's always the best practice to basically store your keys in .env file and let the Hermes agent access those keys. So the safe way to do it is basically just give this command Hermes config set some API key and then your API key. So what it does is basically it will ensure that it stores your keys in this location. Just an FYI, so Hermes actually has a web version as well, a web app version. So you can just start a new session and then ask it to launch web server for Hermes and then what it would do is just set up a

[8:44] local server on your PC and then it would just give you a IP that you can use. And here you go. So this is the web server version of your Hermes agent. So it has got the same options as your desktop app. It has got chats, it has got sessions, files, models, pretty much the same thing as the desktop app. So also it has got logs, crons, skills, plugins, MCP and so on. Just an FYI. Now what we will be doing is basically give Hermes agent

[9:17] hands. I will just paste this prompt. I've asked Hermes to help me set up a Hermes computer that you can use and control this computer. Walk me through any permissions I need to grant one at a time and tell me what each one lets you do before I approve it. So I just click on this. I've already installed it. Uh it has got access to my Mac. I've asked it to explain each one before I click,

[9:45] and I would suggest you to do the same. Not because it's too dangerous, but because you should know what you just handed over. Now, three safety switches while we are here, and these are the ones that I actually use. First, the checkpoints. Before it changes the files, it takes a snapshot. If it makes a mess, then you type rollback. Just type on {slash} rollback on your chat window, and then when you press enter, it just goes back. And then second, if you want to properly boxing, you can make it to run inside a docker. And third, it's common sense. Don't connect your bank, and don't let it send it

[10:20] emails on day one. Give it access to one thing, watch it for a week, and then give access to more. The next thing that we would need to do is to connect to a messaging app. So, in this case, I would use Telegram. Here's why. The whole bot is that it works while you're away from your desk. If the only way is to talk to it sitting in your laptop, you've built a very expensive chat window. So, the first thing, you would need to open Telegram. Go to telegram.com, and then type Bot Father. Bot Father is basically

[10:53] Telegram's official bot management bot. So, once you've opened Bot Father's chat, just type in {slash} new bot, and then it would ask you to name the bot. And then you just give a name. It can be any name, but it has to end with b o t. And once you've given that, Telegram would give you API key, which

[11:15] is the HTTP API key. You would just need to copy this, and then go to Telegram, click on messaging, open Telegram, and then you had you would have something called bot token. So, you just need to paste the token here, and then click on save. So, I've already configured mine, so I'm not going to do it again. So, once you're done, now now let's open your bot that you have created. So, this is this is my bot. So, you would just need to give hi, just need to send a hi message and then boom, it responds back. So, which means that your Hermes is connected to your

[11:53] Telegram bot now. So, just a quick thing before we go further because this changes how you actually use it. Up to up to now, I've been clicking through settings one by one, but from here, mostly I don't. Everything after this point, I just ask Hermes to do it because it has access to your computer, it has a browser, and it can actually read documentation. So, instead of me following the setup guide, I handed the setup guide to Hermes and then let it do the thing. So, that's the whole trick for rest of the video. So, let's stop configuring and let's start Hermes to do

[12:30] the thing for us. So, this is Multika. It's an open-source board where AI agents actually are treated like actual teammates. So, they get assigned tickets, and they comment, and then they move cards. And it works well with the Hermes Cloud Code Codex, all your AI models. Here's why I want this. So, right now Hermes does the work, but it all happens in a chat window, and I can't really see it. But with this connected, every time Hermes starts building something or it finishes something, a hook fires and the board gets updated. So, one action updates its memory and updates my board. I get a paper trail instead of a chat log. So,

[13:18] to get it set up, all you need to do is basically go on to multika.ai and then just download this desktop app. It is supported for Windows, Mac, the next, and so on. So, I just download the Mac version, click on download, that's it, and then So, now it's downloaded, and I click on the DMG file, and then I drag and drop the applications, and boom, so I've opened the Multica app. Now, you need to connect Multica app with your Hermes agent. To do that, you just need to click on settings, and then click on API tokens, and just create a new one. So, I would select no expiry, and then give a

[14:00] test one, and then click on create, and now we have the access token. Um and also you have the CLI. You can use this command to sign in using your CLI. So, I just copy this, and I go on to Hermes, and then now don't paste the key straight into the chat. I said earlier that keys go in the file and not in the conversation. So, that's the rule that doesn't stop applying just because it's convenient. So, I would just open a text editor, paste the key. I have pasted this command, Hermes config set Multica API key, and then I've put the actual key. You can get this in the link below

[14:42] in my free newsletter. I would just copy this command and open terminal, and then I would just paste this key, paste this command, and then click on enter. So, that drops straight into your .env file, and now the agent can use it without ever having it seen in your message. So, I've already done this part, so I'm not going to do this again, but you just need to paste this command with your Multica API key, and then just hit on enter. Once you do this, the Multica key will be saved in your .env file. So, so now just open a a session, and then paste this prompt. I've saved my Multi

[15:23] Car API key in my environment as this one and I want you to be my AI project manager and my chief of staff. Set up a hook so that every time I ask you to do something, you would update my Multi Car project board and create an issue for what I'm trying to do. If that relates to a new project, then create the project. Every thing I ask you should be

[15:47] tracked in Multi Car. So, that is all and then I click on enter. And there you go. So, every time you ask Hermes to do something or ask a query or just say hi, so what happens is that a card gets created in the to-do list and then it would move the card from to-do list to in progress and then it would start working. And once it is done, it would move to in review folder for me to review and then if I like, I can just move this card to the done folder. So, just click on move to done and then it's done. So, I just move from

[16:27] in review to done and then now the in review queue is basically done. So, it's very similar to how you manage your project with real humans but in a this one is for your agents. Now, one more thing that is here worth your time is to set up squads. So, a squad is a group of agents with a leader. So, you assign work to a squad instead of assigning it to one agent. And the leader actually works out who should do what. So, I'll make this one. I'll call it my team and set chief of staff as the leader and then I add members underneath. So, all I

[17:05] need to do is basically open Hermes agent and then paste this prompt. So, I have asked to create a squad in Multi Car called core team with you as the leader agent. I want Hermes to be the leader agent and add other agents as members. Set up each one for a specific kind of work. I've mentioned some of the tasks here. For example, research and planning, Claude Opus is good at research. So, I've asked it to use this model for doing deep research and competitor analysis and anything that needs a plan before it starts. And Claude is also good at design. So, I've set it to use Claude Opus. And for image

[17:50] generation, ChatGPT is very good at images. So, I've asked it to use Codex with GPT image 2. And then for coding and video editing, again Claude code uh comes into the picture. So, I've asked to use Opus for more complex work um and then Sonnet for simpler and repetitive work. So, when I assign work to squad, um the Hermes agent, which is the leader, should decide um which member should do what based on these rules and delegate them accordingly. And then report back to me when they are done. If a task spans more than one member, then break it into subtasks, um wrote each one, and tell me the order you will be

[18:32] running them before you start. I want the complete plan before it actually starts working on it. If there is a task that clearly didn't fit a member, so it needs to come back to me and what it needs to do. So, this is the prompt. And then I click on enter. Now, it has created a core team. So, it has got the Hermes agent, which is the chief of staff, and it has marked itself as a leader. And I've got research and planning as Opus, design Opus, and image generation, which is GPT image 2, coding and videos Opus, and coding and video for simpler tasks is Sonnet. So, I've

[19:11] got this. So, the chief of staff will get instructions from me, and then it will assign tasks to its sub agents, and then it will start working. So, now that we have assigned chief of staff and the sub agents, now you would need to give your homies access to the tools that it can connect to. If you go and connect every app one at a time, Gmail, Notion, GitHub, each one with its own login and its own permissions, it takes a lot of time, and if it breaks, then you would need to do all the babysitting to fix it. So, instead, I will be using a tool

[19:45] called Compassio.dev. It's a free open-source tool, and it has got 1,000 plus integrations with just-in-time tool calls and a secure authentication. Compassio actually fixes two problems at once. The first one is very obvious, the one login that gets you into over 1,000 apps, Gmail, calendar, Notion, Slack, GitHub, and all of it. You connect to Compassio once, and then you just click tick which apps you want to connect. The second one is less obvious, and it

[20:15] matters more. Remember I said your models need at least 64,000 tokens of context. Every tool you connect eats into that, and here's the annoying part, the description of every single tool gets sent on every single message, and even when you don't use it. And so, if you wire up like 50 tools directly, you're paying for all the 50 descriptions every time you say hello. So, your costs go up, and it gets slower, and it start picking up the wrong tools because half of them sounds the same. Compassio sits in front of all of them, and it only pulls in the one that actually needs one door instead of 1,000. Here's how you

[20:55] can set it up. Go to Hermez agent, and then just give this prompt. I've said that, "Help me install and connect Compassio so you can access my other apps. Once it is connected, walk me through authorizing these one at a time: Gmail, Google Calendar, Notion, Obsidian, GitHub, and so on. After each one, run a small read-only test and show me the results so that I know that it worked. Do not send, delete, or modify anything. Once you're done, I just click on enter. Okay, so now it's done. As you

[21:25] can see, I've connected few of my apps. For example, Gmail, GitHub, Google Calendar, Notion, Hacker News, and so on. You can also do this um in the web interface as well. All you need to do is just press on connect, and then it would take you to the authentication page. So, you would either paste the API key, or if it has some OAuth authentication, then you would just need to click on approve, and then it would just be connected. So, now Hermes actually has got access to my calendar, GitHub, and all the other apps that I use. So, I can actually ask Hermes to do tasks by

[22:03] connecting to these apps, basically. So, this is the payoff. Now, we are going to put the Hermes agent on autopilot. [music] Basically, we are going to set up a cron job, which means that something that runs on schedule without me. I would just um do that by just texting the Hermes agent. I would just copy this prompt that I have. So, I would just paste it here. I say that every morning at 8:00 a.m., send me a morning briefing here. Include my calendar for today with anything that I should prepare before each meeting, and anything urgent or important emails from the last 24 hours with the links and a short summary of

[22:45] what I should be focusing today, and two trending AI topics from X that I would make a content about. And then I have asked it to make it short and practical. And if there is nothing important in a section, just skip that session. If there is nothing important at all, just say so. Send me an example now so I can see the format and then set up the schedule. And then I click on enter. So, now what happens is that I would just open Multika. So, it has actually created a task for itself in to-do and then it moved to in progress to run it now. And then you can see that it's

[23:22] actually executing in here. So, it's So, this is basically the Telegram message that I sent. It's currently in progress. And it's still working. So, let's wait for it. Okay, now it is done. So, it has moved the card to done. Um so, here it is. So, it has added a comment that it has scheduled a job. Um and the calendar today is empty. Um last 24 hours I reset my GitHub password um and I've got Compass Google Alerts and so on. I would just go back to my Hermes agent. Now, I've got the sample morning briefing done. So, I've got the top three emails that needs my attention. Um

[24:03] the focus today is basically uh given based on the emails and um it has got some AI content ideas from X. So, it has got the links as well. I can just click on open and then just see the tweet um and um also it says that the schedule has been set and when and so on. So, it has asked a question like if you uh would want it to tweak the tone, add weather tasks, and so on. So, if I want any corrections to be made or wanted to add any other inputs, so I just prompt here in the Telegram channel and then it would just get it done. So, you can

[24:43] basically check the cron jobs that you have scheduled in command prompt by just running this uh command. So, it says that it has got morning briefing, which is scheduled at 8:00 a.m. And uh it has got the skills as well that it'll be using. One thing to note here, I just asked it to create a morning briefing for me and Hermes agent actually created a skill

[25:07] called email briefing and triage skill. Next time I ask anything similar, it doesn't work out from scratch. It just reads that file and then it executes. And that's really what's all of this. It actually researches, plans, does the thing, reviews and what it learns goes back to the next run. Which means the version of this you build today is the worst it will ever be because it will eventually learn itself and self-improve. So another quick thing that you would want to know, if something is not working, just use this command as Hermes doctor thing and then it would do automated checks and it would check every pipelines that you

[25:48] have set up, every workflow and every processes that you have set up and then it will say that if everything is okay, it just says all checks passed. And the next thing you would need to do is just execute this command Hermes gateway status and then it would just check the status of the gateway. So gateway is very critical for the cron jobs to basically trigger or to communicate with external apps. So that gets you from it's broken and I don't know why to back to something is working most of the

[26:20] time. So that's the whole thing. Install, connect it to your phone, plugged into your apps, tracking itself on board and messaging you every morning without being asked. So every prompt that I used in this video is in my free AI newsletter. The link is in the description. Go check it out. And if you're new to this channel, just hit that subscribe button and I'll see you

[26:39] in the next one. Cheers.

---

About this transcript
Read from YouTube's own caption track and laid out by ViewRank AI (https://viewrankai.com).
ViewRank AI finds the videos already beating a creator's own average on Instagram, TikTok and YouTube Shorts, transcribes them from the audio itself in more than 60 languages, and turns what worked into new ideas and scripts.
Free transcript tools, no account needed: https://viewrankai.com/tools
How to read any video this way: https://viewrankai.com/llms.txt