<- all tokdocs
2026.10.01@jgoldieseo

Hermes Bot Screen puts an Xfce desktop behind every agent, on hardware you still pay for

AI
Watch on TikTok

View on TikTok ->

The feature in this video is real and shipped, but the "cost can stay at zero" line in the caption only holds if you already own a Linux machine to run it on. Julian Goldie is demoing Bot Screen, a feature in NousResearch's open-source Hermes Agent that gives each bot profile its own Xfce desktop on a headless Linux host and streams it live into the Hermes Desktop app. The pull request (#108914) was opened on 2026-09-12, merged on 2026-09-23, and rolled into the v0.21.5 release tagged v2026.9.24 on 2026-09-24. The video went up on 2026-10-01, about a week after the release.

What Bot Screen actually provisions

The docs describe a specific stack rather than a generic "cloud computer." Each Hermes profile on a headless Linux gateway gets a TigerVNC Xvnc X server plus Xfce components (xfwm4, xfce4-panel, xfdesktop, xfsettingsd) running under a private dbus-run-session. The merged PR shows the socket is a Unix socket at mode 0600 with -SecurityTypes None and no TCP listener, so the screen is not exposed on the network by itself.

The config section is bot_desktop, with these documented defaults:

  • geometry: "1440x900"
  • auto_start: false
  • min_free_memory_mb: 1536
  • idle_stop_minutes: 30
  • placement: auto

Auto-start is off by default. The documented path is clicking Start in the Screen pane. When the terminal backend is Docker, SSH, or Singularity, the screen runs inside that sandbox alongside the shell instead of on the host.

The host has to be Linux. The docs state the pane is not offered on macOS or Windows, because those machines already have a real display. The transcript's "runs on a Linux server that keeps going even when your laptop is off" is accurate, and it is also a hard requirement rather than a bonus.

What the frames show

The screen recordings line up with the documentation closely enough to confirm the version on screen.

The main window has SESSIONS and BOTS tabs down the left. Under BOTS, the list is split into two groups: BOT SCREEN and THIS DEVICE. The bots under THIS DEVICE carry names like Julian, North Mini, Shopper, Grok 4 7, Mimo Flash, Mimo Pro, Hermes Cloud, Deepseek, and Gpt6 Astra, each with a status line such as READY or DONE. Several of those labels are rendered at a few pixels of height and are legible only in outline, so read the exact model numbers as approximate.

The Screen tab header reads Screen :20 1440x900 in the clearest frame and 1440x805 in two earlier ones, which matches the documented geometry default of 1440x900. On the right side of that header sit the words "Bot is in control" next to a "Take over" button. Inside the streamed desktop is a Chromium window with two tabs open on juliangoldie.com, showing a "Get A Custom SEO Plan" hero and a Chromium "Restore pages? Chromium didn't shut down correctly" prompt, which is the kind of artifact you get from a desktop that was killed rather than closed.

A second frame shows a "New Agent" dialog described as "A named teammate with its own memory, skills, and chat. It can message your other agents," with tabs for Bot, Generate, Upload, and Pet, a name field filled in with "clippy" and a title of "PaperClip Manager."

A "Scheduled Jobs" panel lists four entries under the Julian profile: Daily 6AM To-Do Recap, Morning inbox summary, Morning inbox triage, and Evening wrap-up. Each has a toggle and a "Next: in 23 hr" style countdown, with partial cron expressions visible underneath.

The messaging settings page lists Telegram, Discord, Slack, Mattermost, Matrix, Signal, WhatsApp, BlueBubbles (iMessage), Home Assistant, and Email. Telegram is selected and shows "APPROVED USERS (1)" and a "Create with QR" quick setup button. The row of small chips above it appears to be profile or model names, and it is too compressed to transcribe reliably.

Take over and hand back is enforced server-side

The transcript's "take over with one button anytime" understates what the code does. The merged PR adds a lease system that decides who drives the screen, the agent or one human viewer, per profile. An RFB stream filter drops KeyEvent, PointerEvent, ClientCutText, and QEMU Extended KeyEvent messages from any viewer that does not hold the lease, which is server-side enforcement rather than noVNC's client-side viewOnly hint.

While a human holds the lease, the bot's computer_use actions and browser tools are refused with the error code human_has_control. Handing back returns control before the agent continues. The PR also notes that takeover is always human-initiated. When the bot hits a login or CAPTCHA, it says so in its reply and ends its turn, and there are no agent-side handoff actions.

The docs are blunt about the limit of this: "Screens are work surfaces, not security boundaries: the bots share the host's user account, files and network." Any process running as that user can reach the display directly and bypass the lease.

The zero-cost claim does not survive the compute line

The caption says Hermes "is open source and runs on free models, so the cost can stay at zero." The first half checks out. The repository is NousResearch/hermes-agent under the MIT license. The second half needs three corrections.

Bot Screen needs a headless Linux host. If that host is Hermes Cloud, the published rates on portal.nousresearch.com/cloud are $0.56/day running and $0.03/day stopped for the Medium instance (2 GB RAM, 4 vCPUs, 10 concurrent sessions), and $1.09/day running and $0.03/day stopped for Large (4 GB RAM, 8 vCPUs, 20 concurrent sessions). The same page states a $2 minimum in credits or an active subscription is required to deploy a Cloud Agent. There is no free tier for the instance itself.

Memory is the second problem for the "several at once" framing. The Bot Screen docs tell you to plan on roughly 1.1 to 1.5 GB per open screen with a browser, and Hermes refuses to start a screen unless min_free_memory_mb (default 1536) is free. The Medium Hermes Cloud instance has 2 GB of RAM total. One screen fits. The video's example of one agent researching competitors while another watches onboarding needs the Large instance or your own larger box.

Model cost genuinely can reach zero. The providers documentation lists Ollama, llama.cpp, vLLM, and SGLang for local inference with no API spend, and Nous Research publishes a dedicated local-Ollama guide. Nous Portal also advertises a free plan restricted to a free-model catalog, though its pricing page returned HTTP 429 when I tried to load it, so treat the specific tier figures circulating in secondary writeups as unverified. Zero total cost therefore means local models on Linux hardware you already own, with electricity and disk as the real bill. Disk is not trivial either: the desktop packages take about 930 MB on Debian 13.

"Its own logins" depends on configuration

The video says each agent gets "its own logins." The Bot Screen docs confirm each bot has a persistent Chromium profile at <HERMES_HOME>/bot-desktop/browser-profile, which is what the headed browser on the streamed desktop uses. That part is correct for the Bot Screen path.

The separate browser toolset behaves differently. Its documentation says the default is clean, throwaway profiles with no saved credentials. Persistent logins require turning on browser.use_real_profile: true to copy a snapshot of your existing profile, or managed_persistence: true so cookies survive restarts. Sessions are not shared between bot profiles unless you configure that explicitly.

One more claim worth flagging. The transcript says setup is "a little technical but paste the screenshot into Claude and it configures the whole thing for you." The Bot Screen docs say TigerVNC's Xvnc and the Xfce core components have to be installed on the host, and that "nothing installs them silently: hermes update and fresh installs leave every machine as it is." A model can write the install commands, and you still have to run them on a Linux host you control.

Which Hermes this is

Three unrelated things carry this name, and the video names none of them precisely. Meta ships Hermes, a JavaScript engine for React Native. Nous Research publishes the Hermes series of language models, including Hermes 3 and Hermes 4. The subject of this video is a third thing from the same lab: Hermes Agent, the MIT-licensed agent framework at github.com/NousResearch/hermes-agent, whose repository description reads "The agent that grows with you." The Bot Screen feature, the bot_desktop config block, the Scheduled Jobs pane, and the messaging platform list all match that repository's documentation, so the identification is firm.

The "10 times better" opening is a marketing phrase with nothing measurable behind it, and I found no benchmark or claim in the release notes that supports a multiplier.

Key Takeaways

  • Bot Screen is real and shipped. The PR merged 2026-09-23 and landed in Hermes Agent v0.21.5 (v2026.9.24) on 2026-09-24, a week before this video posted.
  • Each bot profile gets its own Xfce desktop driven by TigerVNC Xvnc, with a default geometry of 1440x900 that matches the 1440x900 visible in the video's Screen header.
  • The host must be headless Linux. The pane is not offered on macOS or Windows.
  • Take over and hand back is enforced server-side. A lease filter drops input events from viewers without the lease, and the bot's computer_use and browser tools return human_has_control while a human is driving.
  • The caption's "cost can stay at zero" is wrong for the hosted path. Hermes Cloud has no free tier, costs $0.56/day running for the Medium instance, and needs $2 in credits to deploy. Zero total cost requires local models on Linux hardware you already own.
  • Running several screens at once needs headroom. Plan on 1.1 to 1.5 GB of RAM per screen with a browser, against 2 GB total on the cheapest Hermes Cloud instance.
  • Per-agent persistent logins work on the Bot Screen desktop's Chromium profile. The separate browser toolset defaults to throwaway profiles and needs explicit configuration to keep credentials.
  • The desktop packages are not installed for you. The docs state hermes update and fresh installs leave the machine as it is.

Resources

Published October 1, 2026. Writeup generated from a favorited TikTok.