Machina

How To Build Your Own Grok Bot For Free (Hermes Full Course)

This is the companion guide to my video on setting up Hermes Agent the way I run it. Keep it open in another window and copy each command, prompt and link as the video reaches it. Every step carries the time in the video where it starts.

On this page

By the end you have Hermes Desktop running on your Mac the way I run mine on a Mac mini at home: Claude as the main model with cheaper fallbacks, my Monograph look, a few bots with their own models, research tools, shared memory, a knowledge base in Obsidian, and a way to text it from your phone.

Last checked 22 November 2026. Prices and screen labels change, so trust what you see on screen when it differs.

What you need

Tools. Prices as of November 2026.

ToolWhat it’s forCost todayAccount
Hermes DesktopThe agent itselfFree and open sourceNone
Nous PortalHermes Cloud, if you don’t run it on your own computerFree, Plus $20, Super $100, Ultra $200 a monthNous Portal
Hetzner CloudA server of your own, the other cloud routeFrom €5.49 a month before VATHetzner
OpenRouterOne key for every modelPay per use, each model’s own priceOpenRouter
Claude pluginClaude models on your Claude planFree, uses your planClaude Pro or Max
MonographMy look and layout for HermesFreeNone
1PasswordKeys and logins your agents can use without seeing them in a chatPaid plans, see its site1Password
last30daysWhat people said online in the last 30 daysFreeNone
ScrapeCreatorsTikTok, Instagram and more for last30days100 free credits, then $47 for 25,000ScrapeCreators
ParallelMain web searchFrom $1 per 1,000 searches, $5 free a month with a card on fileParallel
TinyFishFree web searchSearch and page reading free, up to 500 searches an hourTinyFish
DeepSeek Search APIA third search option$0.75 per 1,000 searchesDeepSeek
Printing PressAgent-ready command-line tools for websites and appsFree, open sourceNone
FirecrawlReading pages search can’t open1,000 free credits a month, top-ups on paid plans from $19 a monthFirecrawl
Browser UseA cloud browser that gets past bot checks$1 free, then $0.06 per browser hour and $5 per GB through home addressesBrowser Use
HonchoMemory shared by all your agents$5 free credit, then $2 per million tokens storedHoncho
ObsidianReading your knowledge baseFree, including for workNone
yt-dlpDownloading videos and transcriptsFreeNone
PhotonA phone number to text Hermes over iMessageFree for up to 10 people, Pro $25 a monthPhoton

A day of use costs what your models and tools spend from their keys. Before the first run you pay nothing if you use a Claude plan you already have, or the few dollars of OpenRouter credit you choose to buy.

How it works

  1. Where it runs

    Your own Mac, kept awake, or Hermes Cloud or a server.

    Which one?

    • Your Mac steps 2 to 4.
    • The cloud the plan or server, then step 4 on.
  2. Models

    An OpenRouter key, a fallback model, and Claude on your plan as the main model.

  3. Look and layout

    Monograph replaces the default look, with Sessions and Bot Mode side by side.

  4. Who you are

    An interview fills Hermes in, and AGENTS.md and SOUL.md hold your rules.

  5. Safety

    1Password holds the keys, and Smart approvals keep Hermes moving.

  6. Bots

    Each bot gets its own job, model, personality, skills and tools.

  7. Tools

    Five kinds, added one at a time:

    • Research last30days, Parallel or TinyFish, and Printing Press.
    • Hard pages Firecrawl and Browser Use.
    • Memory Honcho, shared with your other agents.
    • Knowledge an LLM Wiki in Obsidian.
    • Video yt-dlp.
  8. A team

    Group chats of bots, Hermes as the lead, and Claude Code driven from Hermes.

  9. Your phone

    A new gateway on iMessage through Photon.

Every key goes into Hermes’s settings or your terminal, never into a chat. Every purchase on this page is yours to make.

Build it

  1. 1. Choose where Hermes runs (video 03:19).

    Do you want Hermes on your own computer, or on a machine in the cloud that never sleeps?

    I run it on my Mac mini at home, and the video does too. For the cloud, take a Nous Portal plan: Plus at $20 a month includes one cloud agent with 4 GB of memory and 25 GB of disk, Super at $100 and Ultra at $200 give more use and bigger machines. For a server of your own, use Hetzner Cloud. Either way, follow the rest of this page from step 4.

  2. 2. Install Hermes Desktop (video 04:41).

    On the Hermes Agent site, click the download for your computer. The Mac app needs macOS 12 or later. Open the download, drag Hermes into Applications, open it and click GET STARTED. If your Mac asks to let it install the hermes command, approve it.

    If you prefer the terminal, the site’s one-line install is:

    Command
    curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

    While it installs, know that Hermes can hold several profiles, each a separate setup with its own data. Two businesses can get two profiles, and nothing passes between them.

  3. 3. Keep the Mac awake (video 05:42).

    A sleeping Mac stops Hermes, its bots and its routines. Open the Terminal app (press Cmd+Space, type Terminal, press Enter). The terminal is the text window where you type commands. Paste these one at a time and type your Mac password when asked:

    Command
    sudo pmset -c sleep 0 disksleep 0 displaysleep 10
    Command
    sudo pmset -c womp 1 autorestart 1
    Command
    pmset -g

    The first line stops the Mac and its disk from sleeping on the charger, and lets the screen turn off after 10 minutes. The second lets your network wake it and turns it back on after a power cut. The third shows your settings: sleep should read 0. On a MacBook, closing the lid still sleeps it. A Mac awake all day uses more power, so it’s your call.

    To undo all of it:

    Command
    sudo pmset restoredefaults
  4. 4. Go through the first run (video 06:52).

    When Hermes is ready, click CONTINUE TO SETUP and allow notifications. Type your name, pick an accent colour and click Continue. Under Select a UI mode, choose Advanced, which shows every part of the app, and click Continue. Skip the extra setup it offers next.

  5. 5. Add your OpenRouter key (video 07:24).

    An API key is the long secret code that lets Hermes use your account. OpenRouter gives you one key for hundreds of models, so you can switch models without a new account each time. Create a key on OpenRouter’s keys page and give it a monthly credit limit, so a job that runs away can’t spend more than you chose. Then buy a few dollars of credit. That purchase is yours to make.

    In Hermes, open Settings from the gear at the bottom left. Under Model > Main model, click Change provider, then Other providers, choose OpenRouter and paste the key into the key field there, never into a chat. It should say OPENROUTER CONNECTED. Click BEGIN. We change the model in step 10.

  6. 6. Install Monograph (video 09:44).

    Monograph is my skin for Hermes: a new look and layout that adds files to Hermes’s home and changes nothing in Hermes itself. It needs Node.js 18 or later. Check in Terminal:

    Command
    node -v

    If it’s missing, install it with Homebrew, or get it from nodejs.org:

    Command
    brew install node

    Quit Hermes with Cmd+Q. In Terminal, paste these three lines:

    Command
    git clone https://github.com/machina-exm/monograph.git
    cd monograph
    ./install.sh --full

    --full installs the whole look, with its add-ons, and turns the skin on. If your Mac offers to install git, accept. Open Hermes again. The repository is github.com/machina-exm/monograph.

    Later, to update it, run this in that folder:

    Command
    git pull
    ./install.sh

    To put Hermes back as it was:

    Command
    ./install.sh --uninstall

    To make everything bigger, open Settings > Appearance > Typography and pick a larger UI Scale.

  7. 7. Know the two views (video 11:47).

    The switch at the bottom left of the sidebar swaps the sessions list for Bot Mode, the view of your bots. I use sessions for heavy setup work at the computer, and bots for the day-to-day jobs of each part of the business. If you want only one, Bot Mode does everything. Start with a session for the next step.

  8. 8. Let Hermes interview you (video 12:28).

    Reverse prompting turns the usual chat around: instead of you asking, Hermes asks, until it knows you and your business. Start a new session and paste this:

    Prompt
    Interview me so you can work with me well from today on. Ask one question at a time, wait for my answer, and follow up when an answer is vague.
    Cover, in this order: who I am and what I do, my business (what I sell, to whom, at what price, and how customers find me), my goals for the next 90 days, the work I repeat every week, the tools and accounts I use, how I like you to talk to me (length, tone, structure, emojis or not), how I like to learn new things, what you must always ask me before doing, and what you must never do.
    Ask at least 15 questions and stop when you could explain my business to a stranger.
    Then show me what you will save, split into rules for AGENTS.md, personality for SOUL.md, and facts about me for memory. Wait for my yes before you save anything, and never save a password, a key or a client's personal details.

    Answer in your own words. Say no to anything in its summary that’s wrong before you approve it.

  9. 9. Add a fallback model (video 13:09).

    A fallback model takes over when your main model fails. Open Settings > Model > Fallback models, click + Add fallback, choose OpenRouter and pick a model from a different company than your main one. A model from the same company tends to go down at the same time. I use DeepSeek’s flash model, today ~deepseek/deepseek-flash-latest, so when the main model fails, DeepSeek answers.

  10. 10. Put Claude on your own plan (video 13:45).

    The Claude plugin runs Claude models on your Claude Pro or Max plan, through Claude Code signed in to it. Know two things first: it’s experimental, and it spends your plan’s usage, the same usage your Claude chats draw on. You can install it two ways:

    • From the catalog. Click the + in the top bar, open Capabilities > Plugins, search claude, open claude-subscription-directsdk and click Install, then Install again on the warning.
    • From its link. Send Hermes the plugin’s page and ask it to install it. This works for anything from outside Hermes:
    Prompt
    Install this plugin: https://github.com/NousResearch/hermes-official-plugins/tree/main/claude-subscription-directsdk

    Claude Code must be installed and signed in to your plan. In Terminal, check it with this. If it fails, install Claude Code from Claude Code’s setup page.

    Command
    claude --version

    Then sign in, in the browser that opens:

    Command
    claude auth login

    Forgot this step, as I did in the video at 24:16? Hermes falls back to your fallback model. Send it this, and it opens the sign-in for you:

    Prompt
    fix the claude direct sdk plugin so claude models work

    Then open Settings > Model > Main model, choose Claude Subscription DirectSDK (Experimental) as the provider, pick the newest Sonnet and click Save. In any session, the model picker at the bottom right now shows your Claude models and your OpenRouter models, and you can switch at any time.

  11. 11. Learn your two files (video 16:50).

    • AGENTS.md holds your standing instructions: your context and rules, read at the start of every session, like in Claude Code or Codex.
    • SOUL.md is the personality file: how Hermes behaves and talks, how proactive it is, emojis or not, tables or short sentences. Each bot has its own.

    Facts and rules go in AGENTS.md, and how Hermes talks goes in SOUL.md. To change either, tell Hermes what to change, or ask Hermes where your SOUL.md is and edit the file yourself. In your first days, correct Hermes every time it does something you don’t like, and ask it to save the rule. The model still shows through: some are blunt, some are friendlier.

  12. 12. Add 1Password (video 18:45).

    1Password keeps your keys and logins in vaults that Hermes can use without you pasting them into a chat. The same vault works for Claude Code or any agent you move to later. Install 1Password and its command-line tool from 1Password’s CLI guide. Then paste this:

    Prompt
    Install the official 1Password skill with hermes skills install official/security/1password, then walk me through connecting it to my 1Password app one step at a time. Never show, print or save a secret's value in this chat.

    To have Hermes load its own provider keys from 1Password, run this in Terminal yourself. The 1Password skill’s page has the details.

    Command
    hermes secrets onepassword setup
  13. 13. Set approvals (video 20:30).

    The approval mode decides when Hermes stops to ask you before a command. Click the mode button at the bottom left of the message box:

    ModeWhat it does
    ManualAsks before every action that needs approval
    SmartJudges each action and asks only when it’s unsure
    OffNever asks

    Pick Smart. Manual keeps you approving all day, and the agent stops working on its own. Use Off only on a machine that holds no personal data, cards or logins you’d mind an agent touching. Hermes has defences against prompt injection, instructions hidden in a page or file, but no agent is immune.

  14. 14. Make your first bot (video 22:07).

    Hermes writes its own skills: when you finish a job together, it can save the steps as a skill and reuse them next time. A bot is a separate Hermes with its own job, model, personality, skills and tools. Which job in your week would you hand to a bot first?

    Keep that job narrow. A bot with one job gets a model, skills and tools that fit it, and its chat stays about that one thing. You can right-click in the sidebar and choose New bot, or ask Hermes. In the main chat, I sent this, and you’d swap in your own job and model:

    Prompt
    create a new bot called Youtube that uses deepseek flash model

    To change a bot, click the … next to it and choose Edit to set its icon, colour, name and description. Advanced opens its model, skills, memory, soul and keys. In the bot’s own chat, click its name at the top to open its panel:

    • Model: its provider and model.
    • Reasoning: how hard it thinks. Leave fast models as they are.
    • Cron jobs: its routines, jobs it runs on a schedule. Ask the bot to build them rather than clicking + New.
    • Soul: its own SOUL.md.
    • Skills, Tools, MCP servers: switch each on or off for this bot.

    I made a second bot, Newsletter, with New bot, then set its Provider to OpenRouter and a fast GLM model, today ~z-ai/glm-flash-latest.

  15. 15. Install last30days (video 27:06).

    An outside skill is one someone else wrote. last30days searches Reddit, X, YouTube, TikTok, Instagram, Hacker News, GitHub and the web for what people said in the last 30 days, or 7 days, or 24 hours. An outside skill runs with the same access as Hermes, so have Hermes read it and tell you what it does before it goes in. This prompt does that:

    Prompt
    install the skill at https://github.com/mvanhorn/last30days-skill and, before you install it, read its HERMES_SETUP.md and SKILL.md, tell me what it sends, reads and costs, and wait for my yes.

    Its first run asks you things. It may offer a ScrapeCreators sign-up, which opens in your browser and saves the key itself. ScrapeCreators adds TikTok, Instagram and more, and is cheap. It may ask to read your browser’s cookies for X: decide yourself, it works without them. If a step needs a key, put it in with hermes config set in Terminal, never in the chat. Then run it in Hermes:

    Prompt
    /last30days "your topic"
  16. 16. Set your web search (video 29:09).

    Hermes sets search and page reading separately. Search finds pages, and page reading opens them. This step sets search, and step 18 sets page reading. Pick one:

    • Parallel, which I use. It goes deep. Make a key at platform.parallel.ai, then in Terminal, with your key in place of <key>:
      Command
      hermes config set PARALLEL_API_KEY <key>
      hermes config set web.search_backend parallel
    • TinyFish, the free option. Make a key at tinyfish.ai, then run these in Terminal. The second asks for the key and sets search for you.
      Command
      hermes plugins install tinyfish
      hermes tinyfish setup
    • DeepSeek Search API, the third option. Its answers lean to Chinese sources. It’s in Hermes’s plugin catalog as web-deepseek.

    Pair last30days with Parallel or TinyFish: one searches social media, the other goes deep into the web. Start a new session after any change.

  17. 17. Add Printing Press (video 30:40).

    A CLI, or command-line tool, is a program an agent runs by typing commands. It’s often faster and steadier for an agent than a connector. Printing Press makes agent-ready CLIs for websites and apps, and its public library already has hundreds, 1Password included. Paste this:

    Prompt
    Install the Printing Press library skill with hermes skills install mvanhorn/printing-press-library/skills/printing-press-library. Then, whenever a job needs a website or app I use, check the library for a CLI for it first, and ask me before you install one.
  18. 18. Add Firecrawl for hard pages (video 32:36).

    Search sometimes finds a page or document it can’t read. Firecrawl reads it. Sign up at firecrawl.dev, copy your key from its API keys page, and in Terminal:

    Command
    hermes config set FIRECRAWL_API_KEY <key>
    hermes config set web.extract_backend firecrawl

    The free plan’s 1,000 credits a month stop rather than charge you. Top up a few dollars on a paid plan only if you run out.

  19. 19. Choose a browser (video 33:36).

    Hermes has its own browser. Open it from the surfaces menu at the top right. It offers Link my profile, which brings your logins and cookies. Do that only if you trust it with every site you’re signed in to, otherwise choose Not now and it starts clean.

    I use Browser Use, a cloud browser: a browser that runs on Browser Use’s computers. It gets past many bot checks and suits a Hermes in the cloud. Make an account and an API key at cloud.browser-use.com, then in Hermes open Capabilities > Tools > Browser, pick Browser Use and add the key there. If you rarely need a browser, the built-in one is enough.

  20. 20. Add Honcho memory (video 36:11).

    Long-term memory is what Hermes recalls about you between sessions. Hermes’s own memory works most of the time, and sometimes it forgets things. Honcho keeps memory in one place that Claude Code and other agents can use too, so you never explain yourself twice, even when you change tools. Sign up at app.honcho.dev, then in Terminal:

    Command
    hermes plugins install honcho
    hermes memory setup honcho

    Choose cloud and the browser sign-in, then answer what it asks. Honcho’s Claude Code guide connects Claude Code to the same memory.

  21. 21. Build a knowledge base (video 38:48).

    A wiki is a folder of short linked pages your agent writes from your notes, work and recordings, and reads before it answers. Hermes’s LLM Wiki skill builds it, and Obsidian lets you read it. Install Obsidian, then paste this:

    Prompt
    Set up an LLM Wiki for me with the official llm-wiki skill (official/research/llm-wiki) and the obsidian skill.
    First ask me where the wiki should live, and which folders hold my notes, writing and recordings. Then set WIKI_PATH and OBSIDIAN_VAULT_PATH to that same folder with hermes config set, so both skills use one vault.
    Write the wiki's rules in SCHEMA.md and show them to me before you save. My source files stay where they are: never change, move or copy them. Every page lists the files it came from, and when two sources disagree, keep both and tell me.
    Show me the pages you plan to write, one line each, and wait for my go-ahead. Then build it with helpers in the background, writing index.md and log.md as you go, and tell me when it's ready to open in Obsidian.

    In Obsidian, open the wiki folder as a vault. I rarely open it: Hermes reads it, and Claude Code can read the same folder.

  22. 22. Install yt-dlp (video 40:59).

    yt-dlp downloads videos and transcripts from YouTube and other video sites. In Terminal:

    Command
    brew install yt-dlp

    Then, in your YouTube bot’s chat, build a routine by talking:

    Prompt
    let's build an automation for searching youtube daily

    It asks what to track, what to do with the videos it finds and where the report lands, then sets the schedule. One idea: download each new video from channels you follow and have a vision model such as ~google/gemini-flash-latest say how it was made and what it changes for you.

  23. 23. Build with a team (video 43:04).

    For lots of building, either open many sessions and run them at once, or make a team of bots. I made a Chief of Staff and a UI/UX Designer with New bot, then + > New group chat with both ticked, and renamed it with the pencil to “Building frontend for my SaaS”. A group chat is a room where several bots answer each other. Send it the job:

    Prompt
    I'm building a saas for growing on youtube, what should the UI look like?

    The bot whose field it is answers first, the others reply, and they stop at the room’s round limit with an answer. They know only what you gave them, so give each bot its personality, tools and memory first.

  24. 24. Make Hermes the lead (video 45:29).

    Send every job to your main Hermes on a strong model and let it pass work to the right bot. For example:

    Prompt
    tell the newsletter bot to find the best platform for hosting my emails

    My rules for bots:

    • Create each bot from your main Hermes.
    • Give it a personality, only the tools it needs, and its routines. A 3D bot gets the Blender tools, and the YouTube bot doesn’t.
    • Pick its model by job: Kimi for writing (~moonshotai/kimi-latest), a fast DeepSeek or GLM model for research and light work, and a strong model only for the lead.
    • Ground work in research first: a round of last30days with Parallel or TinyFish. A Research bot on a fast model is a good start.
  25. 25. Delegate and drive Claude Code (video 49:13).

    Hermes can also split a job across subagents, helpers it starts for one errand, each on its own model. The Advanced Delegation plugin lets you name model routes such as fast and deep, or send a helper as one of your bots. Install it in Terminal:

    Command
    hermes plugins install advanced-delegation
    hermes delegation config

    If you already build in Claude Code, keep it and let Hermes drive it. Hermes plans and checks, and Claude Code writes the code. Point Hermes at a project folder and send something like this:

    Prompt
    In my project folder [path], drive the Claude Code CLI to build this: [the job]. Don't write the code yourself. When it's done, check what changed, tell me what it built and anything that failed.
  26. 26. Reach it from your phone (video 52:14).

    The Hermes mobile app is still in development. Until it’s out, set up a new gateway, the part of Hermes that answers on messaging apps. Telegram, Discord, WhatsApp and iMessage all work, and Telegram can give each bot its own chat. For iMessage, Photon lends you a phone number that passes your texts to Hermes. Sign up at app.photon.codes, then paste this:

    Prompt
    let's set up iMessage using Photon. Run hermes photon setup with my phone number, which I'll give you, then set PHOTON_ALLOWED_USERS so only my number gets answers, and restart the gateway.

    Save the line it prints as a contact on your iPhone and text it. The free plan covers up to 10 people.

  27. 27. Tour the settings (video 53:40).

    Most settings you can ask Hermes to change. Each page can apply to one profile or several.

    • Model: main model, fallback models and Subagents.
    • Safety > Approvals: the same modes as step 13.
    • Browser > Browser profile: linking your real browser.
    • Passwords & Logins: Hermes’s own saved logins. I keep mine in 1Password so they move with me.
    • Memory & Context > Persistent memory: MEMORY.md and its Memory Budget. Honcho does more.
    • Memory & Context > Context & compression: when Hermes shrinks a long chat. On a model with a 1 million token window, cap it at 400,000 with the command below.
    • Voice: talking to Hermes. I dictate with Willow instead, the same in every app.
    • Providers > Accounts: OpenRouter, a Hermes Cloud plan’s models, and others.
    Command
    hermes config set compression.threshold_tokens 400000

If something big breaks, run this in Terminal. It checks your setup and names what’s wrong.

Command
hermes doctor

Check it works

  1. Check 1. Claude answers.

    Start a new session, pick your Claude model in the picker and ask it to name the model answering. It should name Claude, with no switch to the fallback.

  2. Check 2. The fallback is set.

    In Terminal, run this. It should show the model from step 9.

    Command
    hermes fallback list
  3. Check 3. The tools are in.

    In Terminal, run these. You should see last30days and Honcho, and pmset -g should still show sleep 0.

    Command
    hermes skills list
    Command
    hermes memory status
    Command
    pmset -g

Did a check fail? Work out which part caused it before you fix anything.

What happenedWhich partWhat to tell Hermes
Replies switch to DeepSeekClaude Code is signed out“fix the claude direct sdk plugin so claude models work”
Hermes looks as it did before MonographThe skin or the restart“Run hermes config get display.skin and tell me what it says.” Then quit and reopen Hermes
Search answers with no sources or from a backupThe search setting“Run hermes config get web and tell me which backend search and page reading use.”
Bots stop overnightThe Mac sleptRerun the step 3 commands and check pmset -g

Keep it running

  • Update Hermes, then check it:
    Command
    hermes update
    hermes doctor
  • After a Hermes update, update Monograph. In its folder:
    Command
    git pull
    ./install.sh
  • Add one tool at a time, start a new session, and test it before the next.
  • Keep every key in 1Password or hermes config set, never in a chat.