Cua's Hermes plugin really does give each run a throwaway machine, but the video's "open source and free" and "97% of native speed" cards do not survive a check against Cua's own sources

A 39.077007 s vertical clip at 720x1280, 30 fps, in an MP4 (mov,mp4,m4a,3gp,3g2,mj2) container carrying HEVC video at 158,618 bps and AAC stereo audio at 44,100 Hz and 64,268 bps, 229,786 bps overall, 1,122,421 bytes on disk, posted at epoch 1792857600 (2026-10-24 16:00:00 UTC) by handle @jgoldieseo, nickname "Julian The AI Expert", over "original sound - Julian The AI Expert", with a 302-character, 51-word description carrying #AIagents #HermesAgent #AItools #OpenSource #Sandbox, and 3,379 views, 151 likes, 2 comments and 24 reposts at capture time (metadata.json mtime 2026-10-25 05:02:56 UTC); I read 14 of the 20 frames extracted at 2 s intervals (001, 002, 003, 005, 007, 009, 011, 012, 013, 014, 015, 016, 017, 019, 020) against 144 transcript words in 8 segments. The frames run a split-screen format: a talking-head square of the presenter at a microphone below, a white explainer panel above. Frame 001 (0 s) shows two app tiles and the caption "NEW HERMES AI"; frame 002 (2 s) reads "HERMES JUST GOT"; frame 003 (4 s) shows a panel captioned "Its own PC" with a badge "Builds a new computer" beside a "Real demo" video card reading "It clicks through the app on each." over two desktops labelled "Linux" and "macOS", with terminal lines "create_space hermes-macos", "macOS VM - Mac mini", "2 Spaces running" and "computer_click "Run an agent here" +2", credited "Video: @trycua"; frame 005 (8 s) shows a koala logo beside the word "Cua" and the badge "A free plugin for Hermes", with the burned-in caption "CUA"; frames 007 through 011 (12 s to 20 s) build a card headed "1 · THROWAWAY COMPUTERS" containing a terminal reading "cua spaces ls" and "space: hermes--" plus the badges "A fresh Space every run" and "Deleted when Hermes finishes"; frame 012 (22 s) catches a mid-animation title glitch rendering "No leftYCV#S"; frame 013 (24 s) resolves it to "No leftovers" with badges "No junk from yesterday" and "No weird old settings"; frame 015 (28 s) reads "Take over any time" with "Open the viewer, grab control"; frames 016 and 017 (30 s, 32 s) show three tiles, "Your laptop", "Another machine" and "Your Cua cloud"; frame 019 (36 s) shows a card "Cua: an agent driving a real desktop" credited "Video: Cua (YouTube)" beside a large "97%" over the words "of native speed"; frame 020 (38 s) shows the GitHub mark, "trycua/cua", "17,400+ stars", "1,100+ forks" and a badge "Open source". The spoken argument is that Hermes gained a free Cua plugin that builds a fresh computer for each job and deletes it afterwards, which removes risky-command confirmations, keeps every run clean, lets two agents run side by side, allows live viewing and manual takeover, runs locally or in the cloud across Linux, macOS and Windows at near-native speed, and is open source and free. The last transcript segment ends at 38.70 s against 39.13 s of decoded audio and 39.077007 s of container duration, so the transcript terminates inside the audio and shows no Whisper overrun; the one transcription error is the product name, which Whisper renders as "Kua" at 7.60 s while frame 005 and the video description both spell it "Cua".

The plugin exists, is Cua's own work, and the chronology runs the way the video implies

At 1.90 s the video says "Hermes just got a free plugin that builds a brand new computer for every job, then throws it away," and at 7.60 s it credits "a free plugin from Kua," which the frames correct to Cua. Both halves hold up. The plugin lives at trycua/hermes-cua-spaces, created 2026-10-20T05:40:42Z, last pushed 2026-10-21T18:56:15Z, MIT-licensed, described by its own repository metadata as "Run Hermes Agent's terminal, file and code tools inside a Cua Space."

The order of events is worth stating because videos in this genre routinely reverse it. Cua wrote the plugin and asked to be listed; Nous Research did the listing. Pull request NousResearch/hermes-agent#142165, opened and merged on 2026-10-20 by teknium1, says so in its own description: "Lists Cua's own Hermes plugin, trycua/hermes-cua-spaces, in the verified tier at the maintainer's request." A second catalog pull request, #142681, opened and merged 2026-10-21, bumped the entry "from 0.1.0 to 0.4.0, at the Cua team's request." The Hermes plugin catalog page for cua-spaces lists maintainer trycua, version 0.4.0, verified tier, last updated Oct 21, 2026. The video posted 2026-10-24 16:00:00 UTC, three days after the 0.4.0 listing. Nothing here is reversed or backdated.

"Builds a brand new computer for every job, then throws it away" is accurate, and the frames quote the real naming scheme

At 2.80 s the transcript describes the create-and-delete cycle, and frames 009 and 011 put the mechanism on screen as a terminal reading "space: hermes--" under the heading "1 · THROWAWAY COMPUTERS". That string is not a prop. The plugin documentation states that with default settings "each Hermes run makes its own Space, named hermes-<task>-<random>, does the work there and deletes it when it finishes." The behaviour is governed by a setting the documentation names explicitly: delete_on_exit, default true, which deletes auto-created Spaces when Hermes exits. The frames' second badge, "Deleted when Hermes finishes", matches that default.

The documentation also records the inverse case the video never mentions. Setting delete_on_exit: false produces a persistent Space called hermes-default that is reused across sessions, and a Space you name yourself, the documentation's example is my-box, "is never deleted." So the throwaway behaviour is the default, not the only mode. The video presents it as the only mode.

"No more clicking yes on every risky command" is true only for the Spaces the plugin creates itself

At 14.76 s the video says "No more clicking yes on every risky command." The plugin documentation grants the premise and then attaches a limit the video drops: "Because the Space is the sandbox, Hermes skips its risky-command prompts for Spaces this plugin creates." The Hermes catalog page states the exception directly: "Named Spaces retain confirmation prompts."

That distinction is the whole safety argument, so losing it matters. The confirmation prompt is suppressed precisely because the machine is disposable. The moment a user follows the persistence path the documentation offers, by naming a Space or setting delete_on_exit: false, the prompts come back, because the thing being protected is no longer about to be destroyed. The catalog entry for the plugin is documented as covering "dangerous-command prompts for user-created versus disposable spaces," which is the same split. A viewer who takes "no more clicking yes" as a property of the plugin rather than of throwaway Spaces will configure a persistent Space and find the prompts still there.

"Linux, Mac, or Windows" is listed as supported, but the capabilities are not equal across the three

At 32.20 s the video says "Run it on your laptop, a spare Mac mini, or the cloud, Linux, Mac, or Windows," over frames 016 and 017 showing tiles for "Your laptop", "Another machine" and "Your Cua cloud". The three operating systems are genuinely listed: the Hermes catalog page describes the plugin as running Hermes tools "inside isolated Linux, macOS, or Windows machines called Cua Spaces."

The parity stops at the terminal. The plugin README restricts the richer tools to one platform: "Browser tools, computer_use, Bot Desktop - Not in a Windows Space ... Not in macOS or Windows Spaces." Pull request trycua/hermes-cua-spaces#3, which added that capability in version 0.3.0, confirms the restriction is deliberate: the implementation returns None with a warning for Windows and macOS Spaces because the underlying display infrastructure cannot run on those platforms. Linux Spaces alone can host Bot Desktop, computer_use and the browser. The same pull request records that the feature depends on a Hermes commit that was merged but unreleased at the time, with v0.22.2 the then-current release, so the capability also carries a Hermes version floor.

The README adds a performance warning that cuts against the video's casual pairing of hardware and guest OS: "A Windows Space on an Apple silicon Mac runs by emulation and is slow; use a Windows Space on an x86 machine or in your cloud account for real work." The video's "a spare Mac mini" and "Windows" sit in the same sentence at 32.20 s. Cua's own documentation says that particular combination is the one not to use.

Version 0.4.0, per pull request #142681, narrowed the gap for Windows by running Hermes tools in Git Bash, matching how Hermes behaves on a Windows PC, and by auto-installing Git for Windows and Python inside a Space from pinned official installers verified by SHA-256. The README notes the prerequisite plainly: "A Windows Space needs Git for Windows."

The "97%" card cannot be traced to any current Cua source

Frame 019 at 36 s displays "97%" over "of native speed" next to a demo card credited "Video: Cua (YouTube)", and at 35.90 s the narration says "it runs almost as fast as the real machine." I could not source the number.

The current trycua/lume README, which covers the Apple Silicon virtualization layer this claim would have to come from, makes no performance claim at all. It describes Lume as a way to "Create and manage local macOS and Linux VMs on Apple Silicon using Apple's Virtualization.Framework" and gives no percentage, no benchmark and no speed comparison. The main trycua/cua README repeats that same one-line description of Lume without a figure.

The one dated, first-party Cua benchmark I could reach does not produce 97% either. Cua's post Apple Silicon and macOS VMs: 11–16× Faster LLM Inference with llama.cpp, published 2026-08-11, reports VM throughput as a percentage of host across three models: TinyLlama 1.1B at 98.25% of host for prompt processing and 72.06% for token generation; Gemma 4 12B at 99.59% and 94.82%; Muse Glimmer 30B with speedups given but no host percentage. The spread runs from 72.06% to 99.59%, and 97% appears nowhere in it. Those figures also measure GPU-passthrough LLM inference inside a VM, which is a different quantity from the general desktop responsiveness a viewer would infer from "almost as fast as the real machine."

A figure of "up to 97% of native speed" does circulate in third-party directory listings and aggregator write-ups of Lume, which is the most likely route by which it reached the video. It is not in the primary sources now, and I am recording it as unverified rather than repeating it.

"Open source and free" is right about the plugin and wrong about Cua Spaces

At 37.80 s the video closes on "open source and free," and frame 020 pairs that phrase with the GitHub mark, "trycua/cua", "17,400+ stars", "1,100+ forks" and an "Open source" badge. Two separate problems sit in that one card.

The licensing claim splits in a way the card flattens. The plugin itself is MIT, and its README says so: MIT, "the same as the Cua SDKs." But the product the plugin depends on is licensed differently. The trycua/cua README states that Cua Spaces is "source-available under FSL-1.1-MIT" and "free to use, self-host and build on." FSL-1.1-MIT is a Functional Source License that converts to MIT on a delay; source-available is not the same as open source, and Cua's own README uses the narrower term for Spaces while using MIT for the SDKs. The video's badge asserts the broader term over a shot of the Spaces integration.

"Free" is accurate for the local path and incomplete for the cloud path the video names in the same breath at 32.20 s. Cua's pricing page lists a Free tier at $0/month that includes "Spaces on every computer you connect", Teleport and sharing, but caps usage at "Up to 2 live sessions and 20 remote hours a month." Pro runs $15/month billed yearly or $18/month billed monthly for up to 5 live sessions and 200 remote hours. Teams runs $25/seat/month yearly or $30/seat/month monthly. The plugin README's own sign-in note draws the same boundary: "sign-in is only needed for Spaces on your cua.ai account; a local Space needs none." The README also carries a time-limited promotion, "Hermes users: get your first month of Cua Spaces Pro or Teams free with code HERMES-3EPX2A (first 50, until 15 Nov)", which is itself evidence that the paid tier is the expected destination for sustained use. A promotion for the first 50 users is not a free product.

The video's claim at 24.68 s that you can "run two agents at once, and each one gets its own computer" interacts with that cap. The mechanism is real: the README documents per-agent Spaces via HERMES_CUA_SPACE=build-a hermes --profile a. But two concurrent agents against cua.ai Spaces is exactly the Free tier ceiling of 2 live sessions, and the 20 remote-hour monthly allowance is consumed by both.

The star and fork counts on screen understate the repository by a wide margin

Frame 020 shows "17,400+ stars" and "1,100+ forks" for trycua/cua. The GitHub API for that repository, fetched 2026-10-25, returns 30,882 stargazers and 2,253 forks, with 1,287 open issues, an MIT license, homepage https://cua.ai, created 2025-01-31T15:02:49Z and last pushed 2026-10-25T05:05:07Z.

A repository does not add roughly 13,400 stars and 1,150 forks in the day between the video's post time and this fetch, so the on-screen card was built from a stale snapshot rather than a live read. I am not asserting the exact value on 2026-10-24, because I have no dated snapshot for that day; the defensible statement is that the card understates the current counts by a large margin and should not be cited as a current figure. The repository's self-description has also moved on from what the video's framing suggests: GitHub now returns "Scale computer-use 2.0 with open-source drivers, cross-OS sandboxes, and benchmarks for training, evaluation, and data generation," and the primary language is Rust.

Key Takeaways

Verified: The plugin is real, free to install and MIT-licensed. It lives at trycua/hermes-cua-spaces, created 2026-10-20, and was listed in Hermes Agent's verified catalog by Nous Research on 2026-10-20, bumped to 0.4.0 on 2026-10-21.
Verified: Cua built the plugin and requested the listing; Nous Research listed it. The video's attribution of the plugin to Cua is correct and the chronology is not reversed.
Verified: The throwaway behaviour matches the documentation. Default settings create a Space named hermes-<task>-<random> per run and delete it on exit via delete_on_exit: true, exactly as the frame 011 terminal card shows.
Verified: Live viewing and manual takeover exist. The documentation describes opening the Cua Spaces app or cua viewer to watch the desktop and "click in the viewer to take over; Hermes keeps working next to you."
Partial correction: The video says "No more clicking yes on every risky command." Hermes skips those prompts only "for Spaces this plugin creates." Named Spaces retain confirmation prompts, per the Hermes catalog page.
Partial correction: The video lists Linux, macOS and Windows as equivalent. Terminal, file and code tools work across all three, but browser tools, computer_use and Bot Desktop are Linux-Space only; pull request #3 returns None with a warning on macOS and Windows.
Correction: The video pairs "a spare Mac mini" with Windows in one sentence. The plugin README says a Windows Space on an Apple silicon Mac "runs by emulation and is slow" and directs users to an x86 machine or cloud account "for real work."
Correction: The closing "open source and free" over a Cua Spaces integration overstates the license. Cua's own README calls Cua Spaces "source-available under FSL-1.1-MIT" and reserves MIT for the SDKs and the plugin.
Correction: "Free" covers local Spaces, which need no sign-in. Spaces on a cua.ai account run on a Free tier capped at 2 live sessions and 20 remote hours per month, with Pro at $15 to $18 per month and Teams at $25 to $30 per seat per month.
Correction: The "17,400+ stars, 1,100+ forks" card is stale. The GitHub API returned 30,882 stars and 2,253 forks on 2026-10-25.
Unverified: The "97% of native speed" card is not traceable to any current Cua primary source. The Lume README makes no performance claim, and Cua's dated August 2026 benchmark reports 72.06% to 99.59% of host depending on model and workload, never 97%.
Clarification: "Run two agents at once" works via HERMES_CUA_SPACE and Hermes profiles, but two concurrent cua.ai Spaces is precisely the Free tier's live-session ceiling.
Context: Whisper transcribed the product name as "Kua" at 7.60 s. Frame 005 and the video description both spell it Cua. The transcript ends at 38.70 s against 39.13 s of audio, so there is no hallucinated tail.

Resources

trycua/hermes-cua-spaces on GitHub establishes the plugin's existence, its MIT license, its 2026-10-20 creation date and its 2026-10-21 last push, plus the repository's own one-line description.
hermes-cua-spaces README establishes the install commands, the space, delete_on_exit, image and on settings, the hermes-<task>-<random> naming scheme, the risky-prompt exemption wording, the Linux-only tool restriction, the Windows-on-Apple-silicon emulation warning, the local-versus-relay data boundary and the HERMES-3EPX2A promotion.
cua-spaces entry, Hermes Agent plugin catalog establishes maintainer trycua, version 0.4.0, verified tier, the Oct 21 2026 update date, the settings table defaults and the statement that named Spaces retain confirmation prompts.
NousResearch/hermes-agent pull request #142165 establishes who moved first: Cua wrote the plugin and requested the verified listing, merged 2026-10-20 by teknium1.
NousResearch/hermes-agent pull request #142681 establishes the 0.4.0 bump on 2026-10-21 and the Git Bash and auto-dependency changes for Windows Spaces.
trycua/hermes-cua-spaces pull request #3 establishes that Bot Desktop, computer_use and the browser are Linux-only by design and carry a Hermes version floor above v0.22.2.
trycua/cua repository API record establishes the live 30,882 stars and 2,253 forks on 2026-10-25, the MIT license, the 2025-01-31 creation date and the current project description.
trycua/cua README establishes that Cua Spaces is source-available under FSL-1.1-MIT while the SDKs are MIT, and gives Lume's one-line description with no performance figure.
trycua/lume README establishes that the current Lume documentation makes no native-speed percentage claim at all.
Apple Silicon and macOS VMs: 11–16× Faster LLM Inference with llama.cpp, Cua, 2026-08-11 establishes the only dated first-party VM performance figures I could reach, 72.06% to 99.59% of host depending on model and workload, with no 97% result.
Cua pricing establishes the Free tier's 2 live sessions and 20 remote hours per month, Pro at $15 to $18 per month, Teams at $25 to $30 per seat per month, and that Cua does not host the computers.