01 / START SMALL
Pick a job you can judge.
I’d rather get useful work out of an agent before trusting it with trading money. A profitable-looking dashboard isn’t proof of a working strategy.
Choose one repetitive job you already understand: turning saved crypto project updates into a daily research brief, sorting research into video ideas, or revising a draft from your edits. Give it a small example and say what a good result looks like. You need to be able to spot when it gets something wrong.
02 / HOW I USE IT
My PC is the interface.
The agent works elsewhere.
I prefer keeping the agent away from my everyday computer, where my private files and logged-in browser accounts live. That’s why I initially used AWS. Now it runs on a separate laptop.
I use the Hermes desktop app on my main PC to give instructions and review results. The laptop runs the agent and its tools, with the work files I provide. The cloud model supplies the reasoning.
Instructions & review
The work runs here
Uses the context sent to it
A remote connection doesn’t automatically give the agent my PC’s files or applications. I deliberately provide the files and access it needs, rather than my entire personal environment. [1]
Keeping it separate reduces exposure compared with giving it broad access on my personal PC, including the scope for accidental deletions or malicious software to affect that PC. It isn’t a security guarantee: shared accounts/files and sensitive context sent to a cloud model still have privacy implications. [2]
03 / USE WHAT YOU HAVE
Choose where the agent will run.
- A spare laptop or computer: use it for the agent if its operating system is supported. Keep your main PC as the interface.
- An existing server: it can fill the same role. You don’t need to rent another one.
- Only your main computer: start with the sample task in your existing AI chat, without computer-control tools. If you want to try a local agent next, agree on limited permissions and copied test files first. A folder by itself doesn’t isolate the agent.
For remote access, use an authenticated private network/VPN; don’t expose the agent’s service directly to the public internet. The prompt below guides you through one suitable route.
04 / GUIDED SETUP
Let your AI walk you through it.
Copy the prompt into the AI you already use, or give it this guide’s link and ask it to help you get started. The prompt also works when your AI can’t open links.
- 01 / YOUR MACHINESWhere will you use it, and where will it run?
Your main computer, any spare computer or existing server, and their operating systems.
- 02 / YOUR BUDGETWhat AI access do you already pay for?
What you have and your extra budget. Zero counts.
- 03 / YOUR FIRST JOBWhat repetitive job do you want help with?
Start with one.
Your AI asks three questions, recommends one route, then guides one step at a time. You approve changes and sign in yourself. A chat assistant can guide you; it can only act on your computer if it has the tools and permission.
Read or manually copy the setup prompt
Copy everything below into a new AI conversation.
Help me give an AI agent one useful job, check the result, and save a repeatable process. Use Hermes as the worked example. Start small, not with automated trading. ASK THESE THREE QUESTIONS TOGETHER, THEN WAIT 1. Which computer will you use to interact with the agent, and where would you like the agent to run? Do you have a spare laptop/computer, an existing server, or only your main computer? Include operating systems where known. 2. What AI access do you already pay for, and what extra budget do you have? 3. What one repetitive job do you want help with? AFTER MY ANSWERS Recommend ONE suitable route and explain it briefly. Do not list every installation method. Ask a follow-up only if needed to proceed. Confirm installation permission and supported operating systems before making changes. Use Adam's arrangement as a reference, not a requirement: Main PC: Hermes desktop app <-> Separate laptop: agent, work files and tools <-> Cloud model. Adam initially used AWS to keep the agent away from his everyday personal computer. Now a separate laptop runs it. The desktop app is his interface for instructions and review; the cloud model provides the reasoning. He deliberately gives the agent the files and access needed for its work. Connecting remotely does not automatically give it access to the main PC's files, browser accounts or applications. Separation reduces exposure, but shared accounts/files and sensitive context sent to a cloud model still have privacy implications. IF I WANT A REMOTE AGENT Do not substitute a local agent on my main PC. Label every action MAIN PC or AGENT MACHINE. Check current official Hermes remote-connection documentation before giving exact instructions. Guide setup of the agent and model access on the agent machine. Start and verify its remote connection service before connecting the desktop on my main PC. Use an authenticated private network/VPN, with service authentication enabled. Do not expose the agent's service directly to the public internet, even if it has a login. Confirm the selected session runs on the intended machine before the first task. Transfer only the sample needed, with my approval. IF I ONLY HAVE MY MAIN COMPUTER Briefly explain the tradeoff. Offer a small trial using fictional or non-sensitive input in my existing chat first, without computer-control tools. If I want an actual local agent, agree on limited permissions and copied test files before setup. A folder alone does not isolate it. Do not require another computer or a rented server; do not call a chat-only trial a completed agent setup. GUIDE ONE STEP AT A TIME Check current official installation, connection and provider instructions; if you cannot browse, ask for the relevant documentation rather than guess. Reuse suitable existing AI access and explain any additional charges. Get approval for installation, access changes or spending. Give one action, its expected result, then wait or verify it before proceeding. If you only have chat access, say so. I enter credentials in official login/configuration screens, never in chat. Keep approvals enabled, explain how to stop a run, and ask before sending, publishing, deleting or overwriting important data. No wallet or exchange connections for this trial. CHECK A SMALL RESULT Once a model reply and the execution machine are confirmed, explicitly create an agent-test folder ON THE MACHINE RUNNING THE AGENT. Use a non-sensitive sample of my chosen task. For this guide's worked example, use "Turn crypto project updates into a daily research brief." Save these three clearly fictional updates in agent-test/updates.txt. All projects and updates are made up, not real news: A · Project A — launch announcement: “We launched a portfolio dashboard.” No date or link supplied. B · Project B — governance post: “Proposal: unlock treasury tokens next month.” Amount and approval status not supplied. C · Project C — promotional post: “Our swap engine is the fastest on the market.” No benchmark or supporting evidence supplied. Task: Read only agent-test/updates.txt. Create a new daily-research-brief.md inside agent-test, leaving the input unchanged. For each update, include: project, what changed, source, confirmed facts versus claims, and questions worth researching. Use “unknown” for missing details. Confirm only what the supplied material establishes; an announcement is not independent proof. Don’t invent facts, dates or links, and don’t give buy/sell recommendations. Show me the saved file and its contents. Short expected output (fictional example, not a verified real-world report): Project A — Changed: dashboard launch announced. Source: A, launch announcement; date/link unknown. Facts vs claims: the note announces a launch; actual availability is unverified. Research: is there a working product or demo? Project B — Changed: treasury unlock proposed. Source: B, governance post; date/link unknown. Facts vs claims: this is a proposal, not a confirmed unlock; approval, amount and exact date unknown. Research: where are the vote and token schedule? Project C — Changed: speed claim published in the supplied post. Source: C, promotional post; date/link unknown. Facts vs claims: the claim exists, but “fastest” is unsupported. Research: are there independent benchmarks and comparable test conditions? Have me open the actual file and compare it with the input. Verify that a proposal has not become an approved unlock, an unsupported performance claim has not become a proven result, and missing details stay "unknown". Check the original input is unchanged. Do not invent facts or give buy/sell recommendations. In a chat-only trial, paste the updates and inspect the reply instead; identify folder and file access as not tested. If a step repeatedly fails, stop and diagnose a redacted error rather than keep retrying. SAVE AND REPEAT After a checked success, save the instructions, checks and approved example as a workflow. Repeat the same format with my own saved announcements and research notes, keeping source references and the same facts-versus-claims checks. Try another sample before expanding access. Finish with how to run it again. OFFICIAL REFERENCES — recheck at setup time Remote desktop connection: https://hermes-agent.nousresearch.com/docs/user-guide/desktop#connecting-to-a-remote-backend Installation: https://hermes-agent.nousresearch.com/docs/getting-started/installation Providers: https://hermes-agent.nousresearch.com/docs/integrations/providers Security details if needed: https://hermes-agent.nousresearch.com/docs/user-guide/security Begin with the three questions only. Wait for my answers.
Optional: what a separate-machine setup involves
Agent machine: install Hermes, connect your chosen model and start its authenticated connection service over that private network/VPN. Keep that service running so the desktop can connect.
Main PC: use the desktop app’s Settings → Gateways → Connection mode → Remote gateway connection to sign in to that machine. Check that your session is using it before starting work. Official remote-connection guide.
The app also supports local sessions. Installing the desktop app alone does not put your agent on another computer. Your helper should check the current instructions for your chosen machines, not send you through a local setup by default.
Optional: installation, access and troubleshooting
Have your helper use the official installation instructions for the chosen machine and check your existing model access before suggesting more spending.
Keep approvals on, give it only the access needed, and learn how to stop a run. If setup fails, share the relevant error with secrets removed. Stop repeated retries and fix the failing step before adding anything else.
05 / CHECK THE WORK
Turn crypto project updates into a daily research brief.
On the machine running the agent, create a folder named agent-test. Save these three clearly fictional updates in agent-test/updates.txt. They are practice material, not real project news.
A · Project A — launch announcement: “We launched a portfolio dashboard.” No date or link supplied.
B · Project B — governance post: “Proposal: unlock treasury tokens next month.” Amount and approval status not supplied.
C · Project C — promotional post: “Our swap engine is the fastest on the market.” No benchmark or supporting evidence supplied.
Ask:
Read only agent-test/updates.txt. Create a new daily-research-brief.md inside agent-test, leaving the input unchanged. For each update, include: project, what changed, source, confirmed facts versus claims, and questions worth researching. Use “unknown” for missing details. Confirm only what the supplied material establishes; an announcement is not independent proof. Don’t invent facts, dates or links, and don’t give buy/sell recommendations. Show me the saved file and its contents.
SHORT EXPECTED OUTPUT — FICTIONAL EXAMPLE
Project A — Changed: dashboard launch announced. Source: A, launch announcement; date/link unknown. Facts vs claims: the note announces a launch; actual availability is unverified. Research: is there a working product or demo?
Project B — Changed: treasury unlock proposed. Source: B, governance post; date/link unknown. Facts vs claims: this is a proposal, not a confirmed unlock; approval, amount and exact date unknown. Research: where are the vote and token schedule?
Project C — Changed: speed claim published in the supplied post. Source: C, promotional post; date/link unknown. Facts vs claims: the claim exists, but “fastest” is unsupported. Research: are there independent benchmarks and comparable test conditions?
Open the file and compare it with the inputs. The proposal must not become an approved unlock, and the speed claim must not become a proven result. Check that missing details stay “unknown,” all three sources are identified, and the input file is unchanged.
Doing the chat-only trial? Paste the three updates and request the brief in the reply instead. Check it the same way; folder and file access haven’t been tested yet.
06 / KEEP WHAT WORKS
Save the process, not just the chat.
Once you’ve checked a good result, ask the agent to save a short workflow: the input it needs, the steps, the output format, checks and things it must not do. Keep one approved example beside it.
Then repeat the same format with your own saved announcements and research notes. Keep their sources, separate facts from claims and leave missing details as “unknown.” Try another sample before expanding the workflow; you can turn the checked process into a skill later.
A couple of practical questions.
Can I use the desktop app without running the work on my main PC?
Yes. That’s how I use it: the desktop app connects to the separate laptop. Select the remote connection, then deliberately provide any files or app access the task needs. It doesn’t make everything on your main PC available automatically.
Do I need to buy hardware or rent a server?
No. My setup is a preference, not an entry requirement. Use what you have; if that’s only your main computer, begin with the limited trial above.
Official sources and scope
My setup is my own experience. Technical references checked 1 November 2026; your helper should recheck them during setup. These are documentation checks, not fresh installations on every operating system.
Start with one job.
Get a result, check it and decide whether it’s worth repeating. For updates on what I’m building, join my free Telegram channel.