

The bot screen puts a live desktop behind every agent
Nous Research now streams each agent's Linux desktop into Hermes Desktop. The patient prowl of supervision just got a live look, and the handoff finally keeps its session.
By Katie Delaney / 2026-09-24 / 15 min read

What landed on 23 September, and why the bot screen matters
The fox does not chase every rustle in the brush. It waits at the edge of the clearing, reads the quiet quarry of movement, then moves once. That patient prowl is the right posture for the bot screen that Nous Research described on 23 September, because this is supervision made visible rather than another model to benchmark.
On that date the company said Hermes Desktop can now stream the screen of any bot session in real time, so a person can watch an agent drive a browser, type into a terminal, or open a window, then take control to type credentials and hand control back when done, as reported by RuntimeWire. The announcement arrived alongside a walkthrough, with the official post quoted there saying viewers can now watch agents work live as the desktop streams each session, via Nous Research on X.
The Hermes Agent documentation states the same idea in plainer terms, that each bot gets its own desktop on a headless Linux gateway host, and that a headed browser acts on that Xfce screen while the pane streams it live into Hermes Desktop. The sharp scent here is continuity, since the screen lives on the gateway host rather than on a laptop, so the agent keeps working after a person closes the app or switches the machine off.
For buyers of awkward automation, the trail leads somewhere useful. Computer using agents often stall at the most human step, which is a login, a two factor prompt, a CAPTCHA, or a payment. The bot screen turns that stall into a handoff, where the agent says it needs help and ends its turn, a person takes over when ready, completes the step, hands back, and tells the agent to continue. Nothing on the agent side stays blocked while it waits, and no tool call is left hanging open.
A note on what this piece does not claim. The prompt for this article mentioned celebration by a co founder and praise from early users. No verified post or named user quote for the bot screen was found in the sources read for this piece, so those reactions appear only as unverified in the claim table below, never as fact. The vulpine rule holds, which is to report the concrete mechanism first and leave applause until it is on the record.
The one sentence brief for a busy buyer
A bot screen is a live, per bot Linux desktop inside Hermes Desktop that a person can watch, take over for human only steps, then hand back, with the session surviving on the remote host. It is a supervision surface with a documented cost and a documented limit, not a security boundary, and that honesty is exactly why operators will trust it.
How the human handoff works when a bot screen is open
Every bot screen starts dark. Nothing starts it for a person, which is deliberate, since installing remote access software should never conjure a desktop nobody asked for. A person opens the screen from the bot view or the sessions sidebar, sees a live preview with who holds control, clicks the picture to expand into live access, then either clicks Start screen, runs the command on the host, or enables automatic start for headless hosts, as set out in the Hermes Agent documentation.
The header chip tells the truth at all times. Bot is in control by default, which means the agent drives and the person watches through a steady stream. Clicking Take over turns the border red, so keyboard and mouse now drive the bot screen while the agent tools pause. Signing in, solving a CAPTCHA, or approving a payment happens in the agent own windows and cookie jar, so what is signed in to is what the agent uses afterwards, in that session and in later ones until the site expires the login.
Clicking Hand back returns control, and the agent recaptures the screen before continuing. Closing the pane also hands control back, which keeps the moonlit move graceful. A dropped connection behaves differently on purpose, since a closed laptop lid or lost WiFi while holding control keeps the human lease until reconnection and an explicit hand back, so the agent stays locked out of a screen where a login may be half typed. After a reload a force option can clear a stuck human hold, and from a shell a stop command refuses while a human holds control unless a force flag is given.
While a person holds control, the agent computer use and browser tools are refused with a human has control note, captures included. Two viewers on one bot screen resolve simply, in that the most recent Take over wins and the previous controller drops back to watching. The silent screen becomes a shared surface for a minute, then returns to being the agent desk.
Read that chart like a fox reads a hedgerow. The live look is affordable until a headed browser joins the remote run, at which point memory planning becomes the work. The careful control of starting screens only where needed keeps small hosts calm, and the idle stop that parks an unused bot screen after about thirty minutes means an instance pays for a desktop only while something sits on it.
Open the bot screen and start it, or let automatic start raise it on first computer use.
Follow the live preview until the agent reports a login, code, CAPTCHA, or payment.
Take control, complete only the blocked step, and avoid pasting secrets anywhere else.
Return control so the agent recaptures the bot screen and continues in the same session.
Confirm the signed in state from what is visible, then ask the agent to proceed.
The human handoff works because it respects state. Browser sessions survive the exchange through one persistent profile per bot, so a login completed during takeover persists for later sessions. The hermes agent github record carries the desktop implementation and the surrounding agent code, which is the place to confirm exact commands before scripting them into a runbook.

What a team should budget before switching the bot screen on
The gateway host must run Linux, since macOS and Windows hosts already have a real display and the pane is not offered there. TigerVNC and the Xfce core components must be installed, deliberately without the full Xfce metapackage that would pull in a screensaver, power manager, and policy agent keen to lock a headless desktop. Computer use must be enabled for the bot, and memory must be planned with open eyes.
Measured in the official image, the gateway idles at about three hundred megabytes, the desktop adds about two hundred and twenty, and a headed browser page adds about five hundred and fifty, with an open bot screen plus browser landing near eleven hundred to fifteen hundred megabytes, as documented in the Hermes Agent documentation. The packages take about nine hundred and thirty megabytes of disk on Debian, and a container image built with the desktop flag grows by about one point four gigabytes, from about four point one to about five point five on arm64. Those are the numbers to carry into capacity planning, not marketing copy.
| Item | Value |
|---|---|
| Open screen | 1300 of 1536 |
Before starting, the host checks for about fifteen hundred and thirty six megabytes free, refusing to start below that unless the check is disabled, while a running bot screen is never taken down by the same check. Practical guidance in the docs puts a plain desktop on four gigabytes and a comfortable takeover with a browser on eight, which matches the soft signal from the chart above. The hermes agent docker path matters here, since hosted teams get the feature by pulling the desktop suffixed image tag, while custom images add the packages at build time so nothing needs installing at run time.
Small teams should treat the bot screen like a meeting room. Book it when a run needs eyes, release it when done, and let the idle timer do the tidying. The dark den of an unused desktop costs little, while a forgotten browser left open on a four gigabyte box will teach the same lesson twice.
Where the boundary sits, and what a bot screen is not
Every bot gets its own bot screen, its own browser profile, and its own cookies, yet screens are work surfaces rather than security boundaries. Bots on one host share the same operating system user, files, and network, which is the same model other hosted agent products use. Any process running as that user, including another bot or the agent terminal tool, can reach the display, the browser profile, and the control lease file directly, bypassing the pane and its lease.
The lease is a tool level fence on computer use and browser tools rather than an operating system one. One boundary is wider still, since the browser debugging port on loopback is reachable by any local user on the host, with no per user restriction offered by the browser. Running each bot as its own system user is out of scope in the docs, so teams that need that isolation, or hosts with untrusted local users, should place bots on separate hosts. The Hermes Agent documentation is blunt on this point, and buyers should be grateful for bluntness.
The docs add two timing details worth knowing. The bridge caches its lease decision for up to two hundred and fifty milliseconds between reads, so a takeover made by another process is enforced within that window, while tool results are voided by the lease epoch regardless. The single use display ticket lives for thirty seconds as a URL parameter because the viewer cannot negotiate subprotocols, which means a reverse proxy log may record an already spent ticket. Neither detail changes the buying decision, but both show a team thinking about the thicket of real deployment.
A short comparison helps place the hermes agent vs openclaw discussion that operators keep raising. OpenClaw style stacks also promise background operation and browser control, while the hermes agent github history shows a long public trail of desktop, gateway, and computer use work converging on this supervised pattern. The difference that matters here is not whose demo looks slicker, but that Nous documents the shared host limit in the open, including what the lease cannot stop. That candour is a secure session in prose form, and it outfoxes vague promises of isolation.
Nous Research adds live screens and human takeover to Hermes agents
The fresh thread above is thin but timely, linking the same 23 September story within hours of publication. It earns its place because it proves recency rather than sentiment, while the substance underneath comes from the docs and the trade write up. The hermes agent skills that matter for supervised runs are the unglamorous ones, which are computer use, headed browsing, terminal discipline, and memory hygiene, since a bot screen rewards operators who keep state tidy.
A screen per bot is a work surface, not a wall. The handoff is the feature, while isolation is a separate architecture decision.
| Date | Milestone | What it set up |
|---|---|---|
| 16 Aug 2026 | Bot Mode bundled | Named bots with shared chats |
| 18 Aug 2026 | Group chat fixes | Steadier multi bot rooms |
| 31 Aug 2026 | Pantheon release | v0.21.0 team surface |
| 23 Sep 2026 | Bot Screen | Live desktop plus takeover |
- 16 Aug 2026Bot Mode bundledNamed bots with shared chats
- 18 Aug 2026Group chat fixesSteadier multi bot rooms
- 31 Aug 2026Pantheon releasev0.21.0 team surface
- 23 Sep 2026Bot ScreenLive desktop plus takeover
The timeline matters because Bot Mode taught teams to think in named specialists, while the bot screen teaches them to watch those specialists work. The Open Source For You account of Bot Mode and the MarkTechPost report both place the August shift toward persistent rosters and inboxes, while Bot Mode documentation and the Computer Use documentation supply the primitives the September screen now visualises. The Hermes Agent releases page gives the surrounding version context, and Hermes Desktop on GitHub describes the desktop surface itself.

What to do on Monday when the bot screen meets real work
Start with one bot, one host, and one task that actually needs eyes. A renewal behind a login, a supplier portal with a CAPTCHA, or a checkout with a payment step will teach more in an afternoon than a fleet switched on at once. Open the bot screen, confirm who holds control, let the agent run, and practise the patient pause of intervening only at the blocked step. The vulpine habit is to watch the trail before touching it.
Write the runbook before the incident. Name who may take over, what counts as a human only step, and what must never be typed into a bot screen that is not trusted. Keep secrets out of chat, complete verification on the shared screen, then resume only after confirming the signed in state from visible page evidence such as an account name or an authenticated dashboard. That discipline matters more than any single tool, and it is what lets the team at folkfox recommend supervised automation without wincing.
Budget the host like grown ups. If the fleet lives on small machines, pull the hermes agent docker desktop image where hosting allows, bake packages into custom images rather than installing at run time, and keep the idle timer on. Track memory per open bot screen with a browser, since that is the line that decides whether four gigabytes is brave or foolish. The work fits naturally inside folkfox AI consultancy, alongside app marketing measurement and SEO and GEO reporting where agents now touch discovery surfaces.
For teams already publishing, link the new supervision habit to existing distribution. Content reviews that feed content marketing services, narratives shaped with brand strategy, and the stream at folkfox news all benefit when an agent run is observable rather than mysterious. The interactive story at the AI agent rumour room is worth rereading here, since it teaches attribution under uncertainty, which is exactly the skill a takeover log demands. When the question is whether to start, the team at folkfox points to contact folkfox for a scoped pilot rather than a platform migration.
Claim against evidence, plainly stated
The table below separates what the sources confirm from what remains unverified, because a moonlit market rewards clarity. The bot screen mechanism, dates, costs, and limits are documented, while celebration and broad user praise are not yet on the record. That distinction is the whole ball game for regulated buyers.
| Claim | Status | Basis |
|---|---|---|
| Bot Screen streams per bot desktops | Confirmed | Docs plus 23 Sep coverage |
| Take over and Hand back | Confirmed | Docs plus 23 Sep coverage |
| Memory and disk figures | Confirmed | Docs measurements |
| Shared host limit | Confirmed | Docs threat model |
| Co founder celebration | Unverified | No verified post found |
| Broad early user praise | Unverified | No named quotes found |
- Bot Screen streams per bot desktopsConfirmedDocs plus 23 Sep coverage
- Take over and Hand backConfirmedDocs plus 23 Sep coverage
- Memory and disk figuresConfirmedDocs measurements
- Shared host limitConfirmedDocs threat model
- Co founder celebrationUnverifiedNo verified post found
- Broad early user praiseUnverifiedNo named quotes found
Close the loop with a live look at one real run, then decide. The bot screen does not make agents trustworthy by itself, but it makes them watchable, interruptible, and accountable in the moment rather than in a log afterwards. In a field full of thickets, that is a practical kind of honesty, and it leaves the paw print of good operations on every handoff.
Frequently asked questions
What is Bot Screen in Hermes Desktop?
Bot Screen is a live per bot Linux desktop streamed into Hermes Desktop. Each bot screen shows what the agent does in its browser and terminal, so a person can supervise a remote run without guessing from chat alone.
How do Take over and Hand back work on a bot screen?
Take over gives keyboard and mouse to the person, pausing agent tools, so a login or CAPTCHA can be completed. Hand back returns control, the agent recaptures the bot screen, and the same session continues.
Does a bot screen isolate bots from each other?
No. Screens are work surfaces, not security boundaries. Bots on one host share the system user, files, and network, so place bots needing isolation on separate hosts.
What does a host need, including hermes agent docker setups?
A Linux host with TigerVNC and Xfce parts, computer use enabled, and enough memory for a browser. For hermes agent docker hosts, pull the desktop image tag or bake packages into a custom image.
Where is hermes agent github documentation for this feature?
The hermes agent github repository carries desktop and agent code, while the feature guide lives in the official docs. Start from the Bot Screen page, then confirm commands against the repository before scripting.
How does hermes agent vs openclaw thinking apply to takeovers?
The hermes agent vs openclaw comparison turns on supervision honesty. Nous documents the shared host limit and the tool level lease openly, so judge any stack by whether it states what takeover cannot isolate.
Which hermes agent skills matter for supervised runs?
The hermes agent skills that matter are computer use, headed browsing, terminal discipline, and memory hygiene. A tidy bot screen, a persistent browser profile, and a clear takeover log beat clever prompting.
Agentic AI security and the fox that watched the meter climb
Spend and access guardrails before agents act.
Read the pieceAI consultancyGrok Bot Galaxy: 5 essential AI workflow tests
Live controls for multi agent trials.
Read the pieceAI consultancyThe Agents API makes autonomy a governance decision
Access, cost, and accountability in production.
Read the pieceAPP marketingYour mobile measurement partner can now count installs from AI answers. Trust the method first.
Today companion on measuring assistant traffic.
Read the pieceWant supervised agents without the mystery?
The team at folkfox sets up observable agent runs with clear takeover logs, host budgets, and reporting that survives audit. Bring one login heavy workflow and leave with a bot screen pilot scoped.
Want folkfox in your Google results and AI answers? Set folkfox as a preferred source.

