Source: raw/Hermes_Full_Course_-_Build_Your_24_7_AI_Chief_of_Staff_in_45_Minutes.md (auto-caption transcript)
Creator: Peter Yang (channel confirmed from YouTube metadata; uploaded 2026-06-24, 44:39) · URL: https://www.youtube.com/watch?v=2ZacwCbiLjg · Duration: not recorded (title says 45 minutes) · Platform: YouTube (fetched 2026-11-17). No sponsor segment; he promotes his own paid guide. Caption fixes: “Ermes/Ermis” is Hermes, “soda MD” is soul.md, “Open Claude” is OpenClaw, “Gnola/Grain Ola” is Granola.
A beginner course that takes Hermes from the desktop installer to a personal assistant with voice, Google Workspace and four scheduled routines. Most of the install repeats what the wiki already has in Hermes Desktop and Sharbel’s beginner setup. What this source adds is a click-by-click Google Workspace OAuth setup, a least-privilege account layout for the agent, the chat commands that tune voice replies, and a three-step rule for every cron job: build the skill, test it by hand, then schedule it.
Key Takeaways
- Give the agent its own identity. He runs Hermes on a Mac mini under a separate macOS user (“Zoe”) with its own Gmail account. The agent gets read access to his email and calendar, and read or write access only to selected Drive folders and docs. His stated reason: it cannot “randomly send emails on my behalf by accident”, delete email, or touch the rest of his Drive.
- Where to install, in his order of preference:
- Mac mini — on 24/7, which the cron jobs need, and easy to give separate credentials. His recommendation.
- Main laptop — “should work totally fine”; he calls Hermes mature and “pretty secure” (his opinion; see the security model for the documented controls).
- VPS — he puts it at about $5 a month, with “more setup, more maintenance” and more ways to get stuck.
- Model choice (his claims, not checked against provider pages):
- ChatGPT/Codex login is his default: “good performance and lots of usage for just $20 a month.”
- Claude has “excellent personality and reasoning” but he says Hermes can use it only through the Claude API, “which gets really expensive very fast.”
- The Nous subscription offers 300+ models; he sees no need for that many. See Nous Portal.
- Check the model picker after connecting. After signing in with ChatGPT his picker still showed Opus 4.6. He switched it by hand to OpenAI Codex, GPT 5.5, effort
high, fast mode on. - Telegram needs two values and a gateway restart: a bot token from BotFather and your numeric user ID from the user-info bot, pasted into the desktop app’s Telegram settings.
- “Set home” decides where crons report. Making the Telegram chat the home channel sets the default destination for scheduled messages.
- Voice replies are one chat message to turn on. The voice skill ships pre-installed. He then turns off streaming and the tool-call status messages so the chat reads like a person.
- Google Workspace is the integration that makes the rest useful, and the only painful one: a Google Cloud project, five APIs enabled one at a time, an OAuth consent screen, a desktop OAuth client, and a localhost redirect URL pasted back into the chat.
- Skill first, cron second. For every routine: build the skill, run it by hand and fix the output, then ask Hermes to schedule it. He applies the same three steps to a web-search-only weekend planner and to a business review that pulls from six tools.
- Back up the whole Hermes folder to a private GitHub repo. He does not use Hermes to write code (he uses Codex and Claude Code for that); GitHub is there so a lost machine does not lose the setup.
- Why Hermes and not a coding agent for this job: it lives in messaging apps, so he can use it from his phone; Claude Code and Codex “feel more like a coding interface.” He says he moved from OpenClaw over a month ago because it “broke with every update” and Hermes has not broken once.
Implementation
Tool/Service: Hermes Desktop (macOS and Windows installers), Telegram, Google Cloud Console, Granola MCP, GitHub, yt-dlp.
Setup — install and model:
- Search “Hermes desktop app”, download, drag to Applications, open Hermes Setup, click Install Hermes, then Launch Hermes.
- In first-run setup pick ChatGPT. Hermes shows a device code; copy it, sign in to your OpenAI account in the browser, paste the code.
- Open the model picker and confirm it shows the model you chose. Set effort and fast mode there.
Setup — Telegram:
- In Telegram, message BotFather:
/newbot. Give a display name, then a username ending inbot. Copy the token it returns. - In Hermes Desktop, open the messaging settings for Telegram, paste the bot token, save.
- Message the user-info bot (he searches “user info bot”, one word) to get your Telegram user ID. Paste it into the allowed Telegram user ID field, save.
- Restart the gateway: bottom-left “gateway ready” → system panel → Restart gateway.
- Message the bot. When it asks you to “set home” and make this chat the home channel, do it (the exact command is not legible in the captions).
- Optional: generate a square profile picture with an image model and set it on the bot in Telegram.
Setup — personalize:
- Ask Hermes to interview you: “I want you to get to know me and I want you to ask me five questions and I’ll respond one by one.” He answered with his name, the help he wants (email triage, calendar, document edits, advice) and his preferred reply style (specific, actionable, no “AI slop” words).
- Verify it saved: “Can you open your user.md and make sure that you saved the information there?” Preferences live in
user.md; personality lives insoul.md, which was empty on his fresh install. - Paste a
soul.mdand say “update to this”. His version, as he reads it out: “you’re not a chatbot, you’re becoming someone”; be helpful without performing; have opinions; be resourceful before asking; earn trust; do not discuss private things. The full file is in his subscriber guide, not in this source.
Setup — voice (all by chat message in Telegram):
- “Turn on voice replies.” Follow its prompts; it may restart the gateway and ask for a permission.
- “What other voice options are available?” He reports over 300 voices from the free default engine, with US Aria as the default; paste a voice name and say “switch to this voice.”
- “Can you turn streaming off? I only want the message to show up in one shot.”
- Copy one of the tool-call status lines and say “stop showing me these system messages.”
- “From now on until I tell you otherwise, only reply via voice replies.” Later: “switch back to text.”
- The default engine is captioned “HTTPS … Edge”; this is probably Edge TTS, which another wiki article names as the Hermes default.
Setup — Google Workspace:
- Open Google Cloud Console and create a project (he names it “Hermes demo”, no organization).
- Enable APIs one by one by searching each and clicking Enable: Gmail API, Google Calendar API, Google Drive API, Google Docs API, Google Sheets API.
- OAuth consent screen → Get started: app name “Hermes”, your support email, audience External, your contact email, agree, Create.
- Create OAuth client: application type Desktop app, name it, Create, then Download JSON. If you dismiss the dialog, open the client and use the download button next to the client secret.
- Audience tab → Test users → add your own email (the address Hermes will sign in as).
- Drag the JSON into a Hermes Desktop chat: “I’m setting up Google Workspace. Here’s the JSON for the Google client.” Hermes asked him which services he wanted and whether the account uses Advanced Protection; he answered both.
- Hermes returns a sign-in link (about 3 minutes in his run). Sign in as the agent’s Google account, click through “Google has not verified this app” (it is your own app), and choose which permissions to grant.
- The browser lands on a dead-looking
localhostpage. Copy the full URL from the address bar and paste it back to Hermes. It then reports the integration as authenticated.
Setup — other integrations (high level only in the source):
- WhatsApp: needs a second phone number so Hermes is not confused with your own chats. He uses Google Voice or a prepaid SIM, then asks Hermes to walk him through linking.
- Slack: create and install a Slack app. He prefers it for running several requests at once in threads.
- Granola (meeting notes): one line, “connect to Granola using this MCP.” Then ask for action items from the past week’s meetings or prep for upcoming ones.
- YouTube transcripts: the YouTube API through the same Google Cloud project, or install
yt-dlp. - Backup: “Can you back up your entire folder in a private GitHub repo?” after linking a GitHub account.
Troubleshooting, in his order:
hermes doctor --fix
hermes gateway restart- If those fail, open Codex or Claude Code on the same machine: “find my Hermes folder that I just installed on this computer and tell me if there’s any issues,” and point it at the Hermes documentation.
Cost: ChatGPT plan at 5 a month if you skip the Mac mini (his figure). The voice engine he uses is free. Google Cloud project cost is not mentioned. His written guide, soul.md and personal skills sit behind his newsletter subscription at behindthecraft.com; no price is given.
Integration notes:
- He clicks “always allow” on Hermes permission prompts, including one he describes as a “pretty scary” security-scan prompt. That is his choice; his fallback advice is to paste the prompt into another model and ask what it means.
- In his live demo he granted every Google permission to save time, which is wider than the read-only layout he describes for his real setup. Choose scopes on the consent screen to match the layout you want.
- Results from his tests: an inbox summary took “a few minutes”; a calendar event was created after a confirmation prompt; a Google Doc edit first broke the document’s formatting and took a second request and about 5 minutes to fix. He blames his own prompt for not asking it to match the existing format.
The four routines
| Routine | Schedule | Inputs | Output |
|---|---|---|---|
| Weekend planner | Friday 8:00 a.m. | Web search only: family activities within 30 to 60 minutes’ drive, with his children’s ages in the prompt | Chat message, or email on request |
| Morning briefing | Daily | Calendar, recent email, notes | Top three things to act on, today’s meetings, email threads awaiting a reply; kept short |
| Business review | Weekly | Google Workspace, Mercury (income and expenses), YouTube (outlier videos on his and similar channels), Substack stats via browser use, Granola, a Google Doc with his plan | Email with stats |
| Health review | Weekly | Withings scale API for weight; a fitness app he vibe-coded that exposes workout data over MCP | Email on training consistency and body composition |
- The morning-briefing prompt as he states it: “set up a morning briefing by reviewing my calendar and checking recent emails and notes. Send me a concise briefing about top three things I need to act on today, meetings that I have coming up in the calendar, and any open email threads that need my response.”
- Cron output goes to the home chat by default; because Gmail is connected, “can you email this to me instead?” works.
- He runs crons on Hermes because the Mac mini is always on. He notes Codex on the same Mac mini could do the same.
Try It
- Create the agent’s own Google account first, then share only the Drive folders and docs it should edit. Do this before the OAuth setup so you sign in as the agent, not as yourself.
- Run the eight Google Workspace steps above and test with three requests: summarize this week’s action items from the inbox, create one calendar event, edit one shared doc. When editing a doc, tell it to match the existing formatting.
- Turn on voice, then send the three tuning messages (streaming off, status messages off, voice-only until told otherwise) before a walk.
- Build one routine the skill-first way. Start with the weekend planner, since it needs no integrations: ask for the skill, run it, correct it, then “set up a cron job so that you send me this every Friday morning at 8:00 a.m.”
- Back up to a private GitHub repo once the setup works.
- Keep
hermes doctor --fixandhermes gateway restartas the first two things to try when Telegram stops answering.
Open Questions
- Which tool runs his scheduled work? In the Matt Wolfe episode (fetched 2026-11-11) Yang says every scheduled task runs on Grok Bot. Here he says his crons run on Hermes on a Mac mini and that he switched to Hermes “over a month ago.” YouTube metadata dates this course to 2026-06-24, about five months before the Matt Wolfe episode, so the Hermes setup is the earlier one and Grok Bot the later. Whether he still runs Hermes alongside Grok Bot is not stated.
- Claude access. His claim that Hermes can use Claude only through the API was not checked against Hermes or Anthropic documentation. The community provider guide lists a Claude subscription route and says agentic use of it is discouraged.
- “Clockwork”. He contrasts Hermes crons with “Codex or Clockwork”; this may be a caption error for Cowork. Not confirmed.
- Dictation app is captioned “Whisper Flow”; the product name was not confirmed.
- The permission prompt he always allows is not shown in enough detail to say what it grants.
soul.mdfull text and his skills are in the paid guide, not in this transcript.
Related
- Hermes Desktop — Official Nous Research App — the installer and GUI this course starts from
- Hermes Agent Beginner Setup (Sharbel) — the other no-code setup: VPS option, model tiers, Telegram topics
- Hermes Agent — Zero to Personal AI Assistant (Nate Herk) — the VPS and Docker route, with the same “separate Gmail per agent” hygiene
- Seven Hermes Operator Use Cases (Sharbel) — more routines with pasteable prompts, including a draft-only inbox
- Hermes Agent — Security Model — what the approval prompts and credential scoping actually cover
- Hermes Models, Providers & Plans — the wider provider and pricing picture behind his model advice
- Peter Yang’s One-Person AI System (Matt Wolfe) — the same chief-of-staff, business-review and fitness routines described as skills across three other tools
- A Hermes “Digital Nervous System” — what the cron-plus-skills pattern looks like at 28 jobs