Four of the five "Hermes upgrades" are not in Hermes Agent, and the "1,300 clicks a day" is one peak day against a 271-a-day average

The artifact is a 72.49 s MP4 (mov,mp4,m4a,3gp,3g2,mj2 container, 4,401,000 bytes, overall 485,694 bps), video stream h264 at 576x1024 and 30/1 fps at 447,083 bps, audio stream AAC stereo at 44,100 Hz and 32,082 bps, format id h264_540p_485687-1; the sidecar audio.mp3 runs 72.568 s. It was posted at epoch 1791797742, which is 2026-10-12T09:35:42Z, by uploader future.with.ai98 (uploader id 7605789824361071623) under the channel name "Future Ai With Julian", credited to "original sound - Future Ai With Julian". The description runs 662 characters and ends with the tags #HermesAgent #AgentOS #NousResearch #AIAgents #VoiceAI #AISEO #AIAutomation. At the capture recorded by the metadata.json file time of 2026-10-27 00:08:33 the counts read 9,021 views, 376 likes, 4 comments, 189 reposts and 492 saves. I read 23 of the 36 two-second frames (1, 3, 5, 7, 9, 10, 12, 13, 14, 16, 18, 20, 21, 22, 24, 26, 27, 28, 30, 31, 33, 35, 36) and then re-extracted the clip at fps=1 into tokdoc_frames_1s/ and read 8 more (7, 17, 66, 67, 69, 70, 71, 72), against a 242-word transcript in 13 segments. The frames show one man speaking to a microphone in a dim room with a screen recording filling the top half. Frame 001 (0 s) carries an "Agent OS" wordmark over a dashboard reading "MONDAY, OCTOBER 12", "16:01:21", "Good afternoon, Julian.", "9 open items and 3 notes in motion.", a prompt box labelled "Ask Hermes anything...", and four counters: "4/5 AGENTS ONLINE", "174 SEO POSTS THIS WEEK", "3 RADAR SIGNALS", "948 THINGS BUILT". Five lower-third cards label the list: "1 · Apollo · talk to your AI" (frames 005 to 008), "2 · Mission Control · every tool, one place" (009 to 015), "3 · Oracle · news → post" (016 to 021), "4 · One memory for every AI" (022 to 028), "5 · Search OS · SEO agent" (030 to 035). A burned-in caption at 1 s frame 007 (6 s) reads "Hermes Apollo", which is how I fixed Whisper's "hermos apollo". Frame 010 shows eight tiles: "Hermes Chat", "Apollo Voice", "Oracle Radar", "SEO Machine", "Video Studio", "Thumbnails", "Builds", "NotebookLM". Frame 016 shows an Oracle card headed "Anthropic ships Claude Managed Agents multi-agent beta" from "@ClaudeDevs" with "10,473 verified likes · 1 source account · post 16h ago", a heat score of 60, a "YOUR ANGLE" block reading "Build 1000-agent workflows in Claude Code; run long-horizon tasks with plan-then-execute", and buttons "Draft this" and "Publish to WP". Frames 022 to 027 show "Memory — Obsidian Vault", a "Memory Galaxy" with "902 memories" and "1,267 connections" and visible nodes "Funnel Strategy", "Julian Goldie" and "Claude", plus a panel reading "One project. Every agent in sync." and the status line "Obsidian configured · 3 projects · append-only history". Frames 030 to 033 show a four-step SEO tool ("Research", "Generate", "Deploy", "History") at "STEP 2 OF 4" headed "One keyword in. Five articles out.", with counters "5 SITES", "0/5 WRITTEN", "0 LIVE", and a Google Search Console panel for aiprofitboardroom.com reading "Clicks per day 1,313" and "Total clicks 46,612" over the axis labels 2026-04-18 and 2026-10-06. The spoken argument is that five add-ons turn a free agent into a full assistant: a voice agent billed to an existing ChatGPT subscription, a single dashboard, a daily news-to-post engine, a memory shared across every AI tool, and an SEO agent credited with growing a site to over 1,300 clicks a day. On the transcript-end check: the last segment is timestamped 68.64 s to 84.56 s while the audio runs 72.568 s, so its end time overshoots the audio by about 12 s, but the words begin inside the audio and the burned-in captions corroborate every one of them ("autopilot" at 1 s frame 071, "bio if you" at 2 s frame 036, "want the full" at 1 s frame 072), so the text is genuine and only the segment boundary is inflated.

The Hermes in question is Nous Research's agent, released 25 February 2026, and it is genuinely free

"Hermes" is an overloaded name. Nous Research has shipped a Hermes model family for years, and its own releases page lists Hermes 3 405B on 24 August 2024, the Hermes 4 Technical Report and Hermes-4-Llama-3.1-405B on 26 August 2025, and Hermes-4.3-Seed-36B on 3 December 2025. The thing this video is about is a different product on the same page: "Hermes Agent", dated 02/25/26 and described by Nous Research as "An autonomous agent that lives on your server, remembers what it learns, and gets more capable the longer it runs."

The repository at github.com/NousResearch/hermes-agent is MIT licensed and carried 257.5k stars when I fetched it. Its README describes "The self-improving AI agent built by Nous Research" with "a built-in learning loop", a terminal UI, messaging gateways for Telegram, Discord, Slack, WhatsApp, Signal and CLI, a cron scheduler, parallel subagent delegation, and backends for local, Docker, SSH, Modal, Daytona and Vercel Sandbox. The official site at hermes-agent.nousresearch.com repeats the same feature set and installs with curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash.

At 0.24 s the video says "these five hermes upgrades are so powerful honestly feels illegal to use them for free". The base agent being free is accurate: MIT license, self-hosted, no license fee. What the line obscures is that running it still costs money for a model provider, and that the five items being demonstrated are not shipped with it.

Four of the five named upgrades do not appear in Hermes Agent

I checked each name against the Nous Research repository README and the official product site. Neither mentions "Apollo", a dashboard, "Mission Control", "Oracle", news scanning, SEO article generation, or memory shared with other AI tools. The community-maintained changelog at hermes-ai.net lists releases back to v2026.3.17 and also contains none of those names; that site's own footer says it is unaffiliated with Nous Research, so I treat it as corroboration rather than a primary source.

The five items are Julian Goldie's own build on top of Hermes Agent. His guide site agentos.guide identifies him as the author, describes the material as "Read-along scripts that pair with the YouTube videos. Built for the AI Profit Boardroom community", and routes the paid tier to a Skool community called AI Profit Boardroom. The dashboard in the frames confirms the authorship from the inside: frame 009 greets "Good afternoon, Julian.", and the memory graph in frame 026 has a node labelled "Julian Goldie".

The one exception is voice. The hermes-ai.net changelog records the Herald release v2026.8.3 as including "Streaming voice with barge-in, wake words, and hands-free control", so a voice capability does exist upstream. The branding does not. "Apollo" is the creator's name for it.

"It runs on the ChatGPT subscription you're already paying for" is true through a third-party repository, not through Hermes Agent

Between 5.44 s and 16.64 s the video says "the first is hermos apollo is a voice agent you can talk to like a real assistant it runs on the chat gpt subscription you're already paying for so you get your own javis without paying a cent extra". Whisper garbled two names here. The burned-in caption at 1 s frame 007 reads "Hermes Apollo", and "javis" is "Jarvis".

The mechanism exists, and it lives in a repository that is neither Nous Research's nor Goldie's. TheSmokeDev/hermes-talk is MIT licensed and its README describes "Realtime duplex voice for your own agent" that "rides your ChatGPT or X Premium subscription" with no API key, using "ChatGPT subscription through codex login" as the default provider with model gpt-realtime-2.1, and an xAI path through hermes auth add xai-oauth with model grok-voice-latest. It also lists GPT-Live, Gemini Live and ElevenLabs cascade voice as alternatives. The string "Apollo" does not appear in it.

The ChatGPT-plan half of that is documented by OpenAI. The openai/codex README says: "Run codex and select Sign in with ChatGPT. We recommend signing into your ChatGPT account to use Codex as part of your Plus, Pro, Business, Edu, or Enterprise plan." What that README does not mention anywhere is voice, audio, or the Realtime API. I tried to fetch OpenAI's help-centre article "Using Codex with your ChatGPT plan" for the authoritative list of what the plan covers and got HTTP 403, so I cannot cite it and cannot confirm from OpenAI whether realtime audio is inside or outside the plan entitlement. The honest statement is that a community tool reuses Codex sign-in credentials to reach a realtime voice model, and OpenAI's published Codex documentation describes that sign-in for coding work only.

A secondary review site, aitoolsreview.co.uk, goes further and claims flatly that Hermes Apollo "does NOT run on ChatGPT Plus/Pro subscriptions" and that paid use requires separate API keys. That contradicts the hermes-talk README. I am citing the repository over the review site because the repository is the thing that implements the behaviour, and I am flagging the conflict rather than resolving it.

"Zero to over 1,300 clicks a day" is the peak day on the chart the video is showing

At 62.48 s the video says "use it to grow our website from zero to over 1300 clicks a day which means more traffic more leads and more customers on autopilot". The Search Console panel on screen at the same moment contains the numbers that undercut it.

Frame 033 (64 s) shows a single-site view headed aiprofitboardroom.com with two tiles: "Clicks per day 1,313" and "Total clicks 46,612". The performance chart below runs from an axis label of 2026-04-18 to 2026-10-06, and the "1,313 clicks / day" tooltip is anchored to a yellow dot at a sharp spike near the right edge, with the line falling away after it. That window is 172 days. 46,612 clicks over 172 days is 271.0 clicks per day on average. The headline figure is 4.85 times the average of the period it is drawn from, and the chart shape shows it as one day, not a run rate.

"Our website" is also doing work in that sentence. Frame 035 and 1 s frame 071 show a six-site view labelled "6 sites · all available data": aiprofitboardroom.com 46,612, aisuccesslabjuliangoldie.com 12,701, aimoneylabjuliangoldie.com 3,162, juliangoldieaiautomation.com 2,935, agentos.guide 1,323, bestaiagentcommunity.com 702, under a banner reading "67,386 clicks" and "4.1M impressions". Those six add to 67,435, which is within 49 of the banner.

Denser sampling mattered here, and this is the specific reason. The 2 s frame 035 catches the same panel mid-animation at "47,454 clicks" and "2.9M impressions", with aimoneylabjuliangoldie.com at 3,158, juliangoldieaiautomation.com at 2,904 and agentos.guide at 1,274. The 1 s frame 069 catches it earlier still at "24,696 clicks" and "1.5M impressions". Any of those three readings would have been a wrong citation. The settled values only appear at 1 s frame 071, two seconds from the end of the clip.

I could not verify "from zero". The left edge of the chart in frame 033 starts low but above the axis, and no earlier data is shown.

Oracle's own cards show the failure mode: a Claude 6 rumour ranked alongside a shipped Anthropic beta

Between 27.20 s and 38.08 s the video says Oracle "every single day it finds the most important news industry ranks it by importance hands you a ready-made hook an angle then one click turns it into a social post or a blog on your website". The frames show exactly that pipeline, including the "Draft this" and "Publish to WP" buttons, and they show what it ranks.

Frame 016 ranks first a card headed "Anthropic ships Claude Managed Agents multi-agent beta" with a heat score of 60. The underlying event is real. Anthropic's own post says "Claude Managed Agents is available today in public beta on the Claude Platform", and the platform docs for multiagent orchestration and workflow runs carry status: beta with the beta header managed-agents-2026-04-01. So the headline is sound.

The angle Oracle attached to it is not. The card's "YOUR ANGLE" reads "Build 1000-agent workflows in Claude Code; run long-horizon tasks with plan-then-execute." The 1,000 figure is real but belongs to a different surface. Anthropic's workflow-runs limits table states the ceiling as "Agents a workflow starts over the run's whole life / 1,000", alongside "Threads working at once in one run / 64", a run lifetime of "24 hours by default", and "Runs open at once in a session / 10 by default". Those limits are documented for Managed Agents workflow runs, not for Claude Code. Oracle moved the number to the wrong product in the one sentence the user is invited to publish.

Frame 018 shows what sits immediately below. Card 02 is headed "Anthropic starts massive training run for Claude 6", sourced to "@TokenGremlin" with 545 likes and a heat score of 48, and its own excerpt labels itself a rumour: "I've been hearing some rumors about Anthropic's next architecture. A massive new training run is reportedly set to begin in the..." Oracle's angle for it is "Track frontier progress; prepare agent stacks for next-gen reasoning models expected mid-2027". Anthropic's live models overview lists the current lineup as Claude Fable 5.1 (claude-fable-5-1), Claude Opus 5.5 (claude-opus-5-5), Claude Sonnet 5.5 (claude-sonnet-5-5) and Claude Haiku 5.5 (claude-haiku-5-5), with legacy entries back through Claude Opus 4.5. No "Claude 6" appears on that page, and neither does the "Claude Next" named in the Oracle Radar list in frame 013. A single 545-like post from an unverified account is scored one notch below a first-party product launch and handed the same "Publish to WP" button.

Card 04 in the same frame reads "OpenAI internal Aeon model solves hard math one-shot" from "@deredleritt3r" with 1,826 likes. There was a real OpenAI math release in that window, reported on 6 and 7 October 2026 and covering several hundred machine-generated results with Lean formalisations from an internal frontier model. I found no first-party source naming that model "Aeon", so the card attaches an unconfirmed name to a real event. Treat it as unverified.

What the frames never show

No frame shows a completed run of the SEO agent. The Search OS screen in frames 030, 031 and 033 is frozen at "STEP 2 OF 4" with "0/5 WRITTEN" and "0 LIVE", and the keyword field holds the placeholder "e.g. hermes mcp server" rather than typed input. The only evidence of the deploy path is a hint line visible at 1 s frame 067: "Each deploy runs: npx @11ty/eleventy then netlify deploy --prod --dir _site. Make sure your sites are linked." That tells you the stack is Eleventy plus Netlify. It does not show an article being written or published.

No frame shows Apollo producing audio. Frame 005 shows a text chat log between "YOU" and "APOLLO" with lines like "At your service, sir. What can I do for you?" and "The console's handling that.", which is a transcript view rather than a demonstration of voice latency.

Two counters on the Mission Control screen in frame 012 read zero while the dashboard is presented as fully operational: "Tracked running goals 0" and "Recorded tokens today 0". The headline counter in frame 001 reads "4/5 AGENTS ONLINE", and the same tile reads "2/5" in frame 009, which is a count-up animation rather than a changing system state.

Key Takeaways

Verified: Hermes Agent is real, free and MIT licensed. Nous Research's releases page dates it 02/25/26 and the repository at github.com/NousResearch/hermes-agent carried 257.5k stars when I fetched it.
Correction: The video presents five "Hermes upgrades". Four of the five (Mission Control, Oracle, the synced memory system, Search OS) appear nowhere in the Nous Research repository README or the official Hermes Agent site. They are the creator's own build on top of the agent. Voice exists upstream; the name "Apollo" does not.
Partial correction: At 11.12 s the video says Apollo "runs on the chat gpt subscription you're already paying for". The mechanism is in a third-party MIT repository, TheSmokeDev/hermes-talk, which rides codex login to reach gpt-realtime-2.1. OpenAI's openai/codex README documents ChatGPT-plan sign-in for Codex and mentions no voice, audio or Realtime API.
Correction: At 62.48 s the video says "grow our website from zero to over 1300 clicks a day". The panel on screen shows 46,612 total clicks for aiprofitboardroom.com between 2026-04-18 and 2026-10-06, which is 172 days and an average of 271.0 clicks per day. The 1,313 figure is a tooltip on one spike near the right edge of the chart, 4.85 times that average.
Clarification: "Our website" is six sites. The final settled Search Console view reads aiprofitboardroom.com 46,612, aisuccesslabjuliangoldie.com 12,701, aimoneylabjuliangoldie.com 3,162, juliangoldieaiautomation.com 2,935, agentos.guide 1,323, bestaiagentcommunity.com 702, totalling 67,386 clicks and 4.1M impressions.
Correction: Oracle's published angle on its top card reads "Build 1000-agent workflows in Claude Code". Anthropic documents the 1,000-agent ceiling for Managed Agents workflow runs, with 64 threads working at once, a 24-hour default run lifetime and 10 open runs per session.
Correction: Oracle card 02 is headed "Anthropic starts massive training run for Claude 6" at heat 48, one notch below a first-party launch. Anthropic's models overview lists Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5 as current, with no Claude 6 and no Claude Next.
Unverified: The on-screen Oracle item claiming "66/70 bugs vs single agent 15-27" does not appear on Anthropic's Claude Managed Agents post or on the workflow-runs and multiagent-orchestration docs pages I fetched. Likewise, no first-party source I reached names OpenAI's internal math model "Aeon".
Context: The hook at 0.24 s is "feels illegal to use them for free", and at 68.64 s the close is "link in the bio if you want the full guide". The guides on agentos.guide are free to read; the site routes paid access to an AI Profit Boardroom Skool community.
Context: Denser frames changed the answer. At 2 s sampling the Search Console banner reads 47,454 clicks and 2.9M impressions, and at 1 s frame 069 it reads 24,696 and 1.5M. Those are count-up animation states. The settled 67,386 and 4.1M only appear at 1 s frame 071.
Clarification: Whisper mangled two names in a 242-word transcript. "hermos apollo" is "Hermes Apollo", confirmed by the burned-in caption at 1 s frame 007, and "javis" is "Jarvis".
Clarification: The final transcript segment is stamped 68.64 s to 84.56 s against 72.568 s of audio, overshooting by about 12 s. The words start inside the audio and match the burned-in captions at 1 s frames 071 and 072, so the text is real and only the end timestamp is wrong.

Resources

Nous Research Releases establishes the 02/25/26 date for Hermes Agent and the separate Hermes model-family releases (Hermes 3 405B, Hermes 4, Hermes-4.3-Seed-36B) that make the name ambiguous.
NousResearch/hermes-agent, GitHub establishes the MIT license, the 257.5k star count at fetch time, and the built-in feature list that contains no Apollo, Mission Control, Oracle, news scanning or SEO generation.
Hermes Agent, Nous Research official site establishes the shipped capabilities and the install command, and corroborates the absence of the five named upgrades.
Hermes Agent changelog, hermes-ai.net establishes the Herald release v2026.8.3 line about "Streaming voice with barge-in, wake words, and hands-free control". Its footer states it is an unofficial community site unaffiliated with Nous Research.
TheSmokeDev/hermes-talk, GitHub establishes the actual mechanism behind the ChatGPT-subscription claim: codex login, model gpt-realtime-2.1, MIT license, no API key, plus the xAI, Gemini and ElevenLabs alternatives.
openai/codex, GitHub establishes the documented "Sign in with ChatGPT" flow for Plus, Pro, Business, Edu and Enterprise plans, and establishes that the README mentions no voice, audio or Realtime API.
Claude Managed Agents: get to production 10x faster, Anthropic establishes the 8 April 2026 public beta and the multi-agent coordination feature behind the Oracle card's top headline.
Workflow runs, Claude Platform Docs establishes the 1,000-agent-per-run ceiling, 64 concurrent threads, 24-hour default run lifetime and 10 open runs per session that I use to correct Oracle's published angle.
Multiagent orchestration, Claude Platform Docs establishes that dynamic workflows are a Managed Agents feature at beta status under header managed-agents-2026-04-01.
Models overview, Claude Platform Docs establishes the live Claude lineup (Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5) and the absence of any Claude 6 or Claude Next.
Julian Goldie Guides, agentos.guide establishes the authorship of the Agent OS layer and the free-guides plus paid-community structure behind the "link in the bio" call.
Hermes Apollo: What Julian Goldie's Voice AI Really Is, AIToolsReview is a secondary review site, cited only to record that it contradicts the hermes-talk README on whether a ChatGPT consumer plan can drive the voice path.

I could not reach OpenAI's help-centre article "Using Codex with your ChatGPT plan" at help.openai.com (HTTP 403), so the authoritative plan entitlement for realtime audio is not cited here.