agentos.guide › blog
The Hermes computer use skill is Hermes Agent's bundled Computer Use skill, which teaches the model to drive your desktop through the built-in computer_use tool.
The honest answer to "skill or tool" is that it's both, because the toolset does the clicking and the skill tells the model how to click well.
Hermes Agent is the free, open-source agent from Nous Research, and it works with any tool-capable model on macOS, Windows and Linux.
This page is the quick start, plus how I'd fit computer use into each profile of an Agent OS.
This is the shortest path to an Agent OS profile that can drive desktop apps, with the safety settings switched on.
| Step | What you do | Command or setting |
|---|---|---|
| 1 | Check the driver that ships with Hermes is installed. | hermes computer-use status |
| 2 | Grant the platform permissions, which on macOS means Accessibility and Screen Recording. | hermes computer-use permissions grant |
| 3 | Switch the Computer Use toolset on for the profile that needs it. | hermes tools, or hermes -t computer_use chat. |
| 4 | Pick the permission mode that fits the job. | computer_use.permission_mode set to standard or bounded. |
| 5 | Give a server bot its own screen if it has no monitor. | hermes computer-use screen start for Bot Screen. |
| 6 | Keep the screen off your own desktop where you can. | A Docker, SSH or Singularity backend, or a Linux Cua Space. |
The rest of this page explains the skill-or-tool question, each step and how I'd map computer use to Agent OS jobs.
Computer use in Hermes is a tool first, and the skill sits on top of it.
| Layer | What it is | What it does |
|---|---|---|
| The tool | The built-in computer_use toolset, running cua-driver. | It takes screenshots, clicks, types, scrolls and drags in the background. |
| The skill | The bundled Computer Use skill, version 2.1.0, at skills/autonomous-ai-agents/computer-use. | It teaches the model the actions, the capture-click-verify workflow and the safety rules. |
| The optional pack | Cua's own skill pack, added with cua-driver skills install. | It adds platform deep dives for macOS, Windows and Linux. |
The skill page lists its authors as Francesco Bonacci (f-trycua) and Hermes Agent, under the MIT licence, for macOS, Windows and Linux.
Francesco Bonacci is also the founder and CTO of Cua, according to Cua's Y Combinator page, and cua-driver comes from Cua's trycua/cua project.
The skill can't click anything without the tool, so switching computer use on really means enabling the toolset.
Hermes drives your desktop in the background on macOS, Windows and Linux.
The docs say your cursor doesn't move, your keyboard focus doesn't change and your virtual desktops don't switch.
It works with any tool-capable model, including Claude, GPT, Gemini and local models on an OpenAI-compatible endpoint.
The old way of desktop automation grabbed your mouse, so you couldn't work while it ran.
The new way lets you and the agent co-work on the same machine.
The docs say the driver ships with Hermes, because cua-driver is pinned and installed by default by the installers, hermes pm install and hermes update.
hermes computer-use status hermes computer-use doctor hermes -t computer_use chat
If the driver is missing, hermes computer-use install installs it, or you can pick Computer Use in hermes tools.
On macOS, grant Accessibility and Screen Recording to CuaDriver, and hermes computer-use permissions grant asks macOS for both.
On Linux you need a reachable display server with AT-SPI on, and Windows needs nothing at install time.
You can also add computer_use to your enabled toolsets in ~/.hermes/config.yaml.
Since v0.22.0, the release notes say computer_use sits behind tool_search by default and loads on demand, which keeps each profile's prompt lighter.
🔥 Want the exact Hermes Agent setup I run inside my Agent OS?
Inside the AI Profit Boardroom you get the installable Agent OS, step-by-step video tutorials, four coaching calls a week and 3,400+ members building real automations.
The skill teaches one habit above all others, which is capture first, click by element number and verify.
computer_use(action="capture", mode="som", app="Mail") computer_use(action="click", element=7, capture_after=True)
The som capture returns a screenshot with a numbered element list, vision returns a plain screenshot and ax returns the list with no image.
Element numbers only last until the next capture, and a stale number is refused rather than clicking the wrong thing.
When a click doesn't land, the skill climbs a ladder from a background element click to a pixel click and then a foreground click with its own approval.
For web page content, the skill points the model at the separate browser tools, because they are more reliable for pages.
On Telegram, Discord or Desktop you can ask "Send me a screenshot of my screen" and get the real image, which suits an Agent OS that lives in chat.
The default standard mode keeps your normal Hermes approvals for every click and keystroke.
For scheduled jobs, the docs describe bounded mode with a capability manifest you review once at launch.
computer_use: permission_mode: bounded capability_manifest: ~/.hermes/cua-manifest.yaml
Anything outside the manifest fails closed, and a missing manifest fails loudly at session start.
--yolo, /yolo or approvals.mode: off launch an unrestricted runtime, and the docs warn that it doesn't protect against prompt injection.
Hard blocks still apply, including empty trash, log out, lock screen and typed patterns such as sudo rm -rf /.
Where nobody can approve, such as a cron job or a headless run, actions are refused rather than auto-approved.
A bot on a headless Linux gateway has no monitor, so v0.22.0 added Bot Screen.
Each bot gets its own Xfce desktop over TigerVNC, streamed live into Hermes Desktop, and the CLI help calls the same feature Bot Desktop.
You click Take over for a login, 2FA, CAPTCHA or payment step, then Hand back, and the bot carries on with your session.
It's off by default, and you start it from the Screen pane or with hermes computer-use screen start.
Every profile has its own screen, browser profile and cookies, which maps neatly onto one bot per Agent OS profile.
The docs plan on about 1.1 to 1.5 GB of memory per open screen with a browser.
If bots are new to you, start with my Hermes Agent bot mode guide.
The docs say computer_use acts on the same machine the bot's screen lives on.
With a Docker, SSH or Singularity backend, the driver runs inside the sandbox on its own display, so it can only touch what the terminal can.
Modal, Daytona and Vercel sandboxes can't host a display yet, so computer_use refuses there unless bot_desktop.placement: gateway opts into the host.
The newest option is Cua's cua-spaces plugin, which the catalog says lets Bot Desktop, computer_use and the browser follow the terminal into a Cua Space on Hermes 0.22.3 and later.
That works in Linux Spaces with the cua command at 0.7.0 or newer and the desktop stack installed in the Space.
That's the update in my video at the top, which walks through Cua's demo footage rather than my own test.
The full setup is in the sibling post on Hermes Cua, and the backends are covered in my Hermes Agent sandbox quick start.
🔥 Want the exact Hermes Agent setup I run inside my Agent OS?
Inside the AI Profit Boardroom you get the installable Agent OS, step-by-step video tutorials, four coaching calls a week and 3,400+ members building real automations.
| Agent OS job | Computer use setup I would use | Why it fits |
|---|---|---|
| An inbox or file-manager helper on your own Mac | The toolset in standard mode. | Every click and keystroke goes through your normal approvals. |
| A scheduled job against a logged-in app | bounded mode with a reviewed manifest. | You approve the scope once, and anything outside it fails closed. |
| A Telegram bot on a Linux server | Bot Screen for that profile. | The bot has its own desktop, browser profile and cookies, and you take over for logins. |
| A risky job you don't want near your files | A Docker, SSH or Singularity backend, or a Linux Cua Space. | The driver runs inside the sandbox on its own display. |
| A text-only or cheap model | mode="ax" or an auxiliary vision model. | The agent can still work from the accessibility tree. |
| Several agents at once | One profile per agent. | Each run gets its own session cursor, so you can see who is acting. |
I'd roll it out one profile at a time, starting with a low-risk helper in standard mode on a machine you can watch.
Nous's compatibility table rates Anthropic's Claude Sonnet and Opus models best overall, with raw coordinates and the numbered list both supported.
OpenAI, Gemini, OpenRouter vision models and local vision models on vLLM, LM Studio or Ollama also work.
Text-only models work in a degraded mode with mode="ax", or Hermes routes screenshots through the auxiliary vision model set in auxiliary.vision.
The docs say a 20-action session on a 1568×900 display typically costs about 30,000 tokens of screenshot context rather than about 600,000.
hermes computer-use doctor first, because it prints a check for every part of the setup.hermes computer-use install or enable the toolset in hermes tools.DISPLAY is set or that you have an XWayland bridge.tccutil and run hermes computer-use permissions grant again.type action won't enter passwords, so use your system's password autofill.For the wider list of add-ons I'd look at next, see the best Hermes Agent tools.
If you want to talk through where desktop agents fit your SEO business, you can book a free SEO strategy session with my team.
It is both. The computer_use toolset performs the actions through cua-driver, and the bundled Computer Use skill, version 2.1.0, teaches the model the actions, the workflow and the safety rules.
Check the driver with hermes computer-use status, grant any platform permissions, then enable the toolset in hermes tools or start a session with hermes -t computer_use chat.
Usually not. The docs say it is pinned and installed by default by the installers, hermes pm install and hermes update, and hermes computer-use install repairs it if it is missing.
Any tool-capable model works. Nous rates Claude Sonnet and Opus best overall, and text-only models work in a degraded accessibility-tree mode.
Yes. With Bot Screen on a Linux gateway, every Hermes profile has its own screen, browser profile and cookies, and you can take over for logins and hand back.
Input goes through the same approval gate as dangerous shell commands, risky key combos are hard-blocked and actions are refused where nobody can approve. A sandbox backend or a Linux Cua Space keeps the screen off your own desktop.
Switch the toolset on, let the bundled skill guide the model, keep standard or bounded mode and move risky jobs into a sandbox.
If you run an agency and want help planning agent work for clients, you can book a free strategy session with my team.
If you want the whole Agent OS with the Hermes tab, the profiles and the training already set up, it lives inside the AI Profit Boardroom, with 3,400+ members and four coaching calls a week.
That is how I'd put the Hermes computer use skill to work.
About Julian
I'm Julian Goldie, an AI entrepreneur, SEO expert and the founder of the AI Profit Boardroom, which has 3,400+ members.
I help business owners scale with AI agents, automation and SEO.
My YouTube channel has 400,000+ subscribers, and I run Goldie Agency, a seven-figure SEO agency.
→ Hermes Agent Sandbox: The Quick Start
→ Hermes Agent v0.22: The Quick Start
The installable system behind every guide on this site — dashboard, agents, memory and pipelines — updated as new tools land, with 3,400+ members and four coaching calls a week.
Join the AI Profit Boardroom →📺 Video notes + links to the tools 👉
🎥 Learn how I make these videos 👉