I Built JARVIS from Iron Man with Claude Fable 5 (INSANE Results!)

At a glance

Zubair Trabzada spends roughly the first seven minutes of this video demonstrating a working voice assistant he calls Jarvis, and the remaining five explaining why he built it and where to get the prompts. The assistant runs as a side panel next to a 3D visualization of his own files and folders, listens hands free for the wake word, answers in a clipped British butler voice, pulls documents and their linked notes out of his vault, searches the web live and returns the result as a card, reads a shared browser tab off his screen in fine detail, and swaps its own underlying model on a spoken command because it holds his OpenRouter key. He says the build took Claude Fable 5, Claude Code, many hours of back and forth, and millions of tokens, and he is unusually direct that the version on screen is tailored to his own files and is not what a viewer will end up with. What is free is a prompt pack inside his free Skool community; the finished V3 build is a paid product. Structurally the video is a demo plus a distribution funnel rather than a build tutorial: no command, no config file, and no line of code appears on screen at any point, so the reconstruction below reads the architecture off the behavior.

The system, in one picture

Before walking the demo, here is the whole machine as the twelve minutes reveal it. Everything in amber or blue was shown working on camera. Everything in the neutral boxes he described out loud but did not demonstrate, or is a role the system obviously fills without the video ever naming the component that fills it.

The cold open: proof first, explanation later (0:00)

The video opens with no title card and no preamble. It opens with the thing working, cut into five exchanges over about twenty five seconds, and the persona lands before the product does.

He starts by making the assistant address the audience directly.

"Hey Jarvis. So, my YouTube audience is watching you right now. Can you give me a quick introduction to everybody?"

"Ladies and gentlemen, allow me to introduce myself. I am Jarvis, Sir Zubair's second brain, chief note wrangler, and the only butler he trusts to remember what he forgot he was building."

There is laughter on the audio track. Then a second exchange that exists purely to show the character has range:

"Hey Jarvis, look alive. We've got a lot of work to do today."

"I am, as ever, disturbingly alive, sir. Point me at the first item on the ledger and let us dispense with the pleasantries."

Then the montage starts doing work, because the next three exchanges are each a different capability, compressed to a single beat.

Retrieval. "Okay, can you pull up this free Skool email funnel information, please?" Answer: "I've pulled up all the information, sir. You can read it yourself, or if you'd like, say read it out loud and I shall." That last clause is worth catching. The assistant volunteers a follow up command it accepts. It is advertising its own interface.

Web research. "Actually, I'm trying to see if there's an alternative to Kit, which is an email marketing platform. Can you search online to see if you can find a cheaper alternative?" Answer: "Cheaper options abound, sir. MailerLite at $7 a month is the cheapest full featured Kit alternative, less than a fifth of Kit's $39 a month Creator plan starting price." That is a real comparison with two hard numbers and a computed ratio in one sentence.

Vision. "I've shared my screen with you. Can you tell me what you're looking at right now?" Answer: "A star history dashboard for Nous Research's Hermes Agent repo. 208.4k stars, 37.9k forks."

Five exchanges, four capabilities, and a character. It is an efficient cold open, and it is also the only part of the video where you see the assistant answer a question about a document, the web, and the screen back to back with nothing cut between them.

The claim, and the price he paid for it (0:29)

Then he steps in front of it and states the thesis.

"So, I officially built Jarvis as my AI assistant with the Claude Fable 5 model. After hours of going back and forth and wasting millions of tokens, this is at a place where it's truly an AI assistant."

Read that sentence twice, because the verb is unusual. He does not say he spent millions of tokens. He says he wasted them. Later, at the build chapter, he repeats the figure without the editorializing, but the first framing is the honest one: getting this to work involved a large amount of output that did not survive.

He promises two things for the twelve minutes. First the full demo of what the thing can do. Second, how you can build something similar. Both promises get kept, though the second one gets kept in a way that will matter later: "something similar" turns out to be load bearing.

The second brain: files and folders as orbiting objects (0:49)

The screen behind him is not a chat window. It is a 3D scene of drifting objects, and he explains what they are:

"What you're seeing right now on the screen, these kind of orbits or these 3D objects, these are all folders and files in my computer. This is basically my AI second brain, but now I have the ability to interact with all of this with Jarvis through my voice."

This is the single most important structural fact in the video and it is easy to skate past. Jarvis is not the brain. The brain is the file system visualization, an existing thing he had already built. Jarvis is a voice layer bolted onto its side. He confirms this later without meaning to, when he names the paid product: "V3 3D Brain with Jarvis." Two components, joined by the word with. If you are trying to reproduce this, you are reproducing two projects, not one, and the video only demonstrates the seam between them.

The ear button and the wake word (1:16)

The controls live along the bottom of the assistant panel, and he introduces the first one:

"If I click on the ear button right here in the bottom, this basically gives it hands free access. So as soon as I say Jarvis, Jarvis will be able to get activated."

The confirmation comes back immediately and in character:

"Hands free engaged, sir. Say Jarvis and I shall answer."

So there are two listening modes, and hands free is opt in. That is a design decision with real consequences: an always on microphone waiting for a wake word is a different cost, privacy, and latency profile from a push to talk button, and he made the always on mode a deliberate toggle rather than the default state.

With hands free engaged he re runs the introduction request live, on camera, and this is where the demo accidentally produces its most interesting result. The wording of the request is almost identical to the cold open. The answer is not. It opens the same way, then diverges:

"Ladies and gentlemen, allow me to introduce myself. I am Jarvis, Sir Zubair's second brain, chief note wrangler... He makes the tools and the videos. I make sure he doesn't lose his own train of thought on camera, which is a full time position."

Same setup, different punchline. The persona is not a canned recording and it is not a fixed script. It is being generated fresh on every turn from a character definition, which is why two runs of the same prompt produce two different jokes about the same relationship. That is a small thing to notice and a large thing to understand about how the system is built.

The thinking state, and what happens inside it (2:14)

He points at the panel while it works:

"He is listening to me the whole time, and every time I interact with him or every time I say something, he says thinking, meaning that, depending on the task that I'm giving him, he's capable of actually going through searching the internet. He can pull up all my notes. He can actually see my screen and everything else."

So there is a visible state machine with at least three states: listening, thinking, speaking. And the thinking state is not decoration. It is the window in which tool selection and tool execution happen, which is why its duration varies with the task. A vault lookup is fast. A web search is not. Exposing that as a labeled state rather than a spinner is the correct choice, because it tells the user that the delay is work rather than lag.

The first real task: pull the email funnel (2:47)

The first live task is retrieval, and he fumbles the sentence:

"Hey Jarvis, can you pull up my um uh school community free school community email funnel information?"

He says "school" where he means Skool, the community platform his free and paid communities live on, and he restarts the phrase mid sentence. It resolves anyway. That is a genuine robustness data point that no scripted demo would ever produce on purpose: the retrieval survived a disfluent, self correcting, partially wrong query.

What comes back is not a single file:

"It's capable of pulling up all of my information for whatever task I'm giving it to. So this is all of my email funnel and all of the connections and all the notes that are related to that particular task."

All the connections and all the related notes. That is graph aware retrieval across a linked vault, not a filename match and an open command. Whatever indexes the second brain understands that a document has neighbors, and the assistant returns the neighborhood.

Live internet research by voice (3:23)

The second live task is the Kit question, asked in full and with the motive attached:

"Hey Jarvis, can you do a search online and find an alternative to Kit that I'm using right now, which is costing me a lot of money?"

"On the wires now, sir. One moment."

While it works he narrates the mechanism, including the output surface:

"It's researching, so it's going to reach out to the internet cuz Jarvis has access to the internet, and it's able to find out all those details. And once it grabs that information, it's going to bring back and show me that in a card."

The answer:

"Right then, sir. Here are the cheaper contenders. EmailOctopus claims to run up to 50% cheaper than Kit, and Brevo offers the lowest entry pricing because it charges per email sent rather than per subscriber."

Then the card, and the card is the part worth studying:

"He has access to the internet, so you can see it pulled up all of that information and also gives me like a little card right here, so that way I can keep this here as kind of my note if I want to. But I can click on Got it and it will move on to the next thing."

The output is dual. The spoken answer is ephemeral. The card is a durable artifact with an explicit dismiss action, and he describes keeping it as a note, which means the card can graduate into the vault. That closes a loop most voice assistants leave open: research done by voice produces something you still have tomorrow.

Now compare the two answers to the same question, because the video gives you both and never mentions the discrepancy.

That divergence is not a flaw in the assistant so much as the nature of the capability. Live web research is a function of whatever the search returned in that minute, and a model summarizing search results will pick a different lead depending on what came back. If you build this, expect the same question to answer differently on Tuesday than it did on Monday, and expect the framing to move too: one run optimized for the lowest sticker price, the other for the lowest entry cost under a different billing model. Both answers are defensible. Neither is stable.

Screen sharing and AI vision (4:32)

The third live capability is the one that would be hardest to fake and is demonstrated most convincingly.

"I can give him access to my screen. So if I click on the screen button on the bottom, I can share any of these tabs. So let's say I want to share this tab with him. Share this screen. Click on share."

"Seeing your screen now, sir. Ask me to look anytime."

Catch the phrasing on the confirmation, because it describes the design. It does not say "watching your screen." It says ask me to look anytime. The grant is persistent; the capture is on demand. That is the right architecture for both cost and privacy: continuously streaming frames into a vision model would be expensive and mostly pointless, and holding a standing permission that you spend only when asked is cheaper and easier to reason about.

Then the question, phrased to make the test unambiguous:

"Hey Jarvis, can you tell me exactly what you're seeing on my screen right now that I've shared with you?"

And the answer, which is the technical high point of the whole twelve minutes:

"A starhistory.com dashboard for Nous Research Hermes Agent, sir. A Python MIT repo. Tagline: the agent that grows with you. Sitting at 21 global rank with 208.4k stars, 37.9k forks, and 1.8k contributors. This week it's pulled plus 434 pushes, and 60..."

Count what that reading actually contains. A site identification. A repository name. A primary language. A license. A verbatim tagline. A global ranking position. Three separate counts with decimal precision. A weekly delta. And it was mid way through a second delta when he cut it off. This is not "I see a GitHub page." It is fine grained text extraction plus layout comprehension, correctly attributing each number to the right label on a chart heavy dashboard.

The subject is a nice accident for anyone who follows this catalog. Hermes Agent is Nous Research's personal agent harness, the same tool covered here in NetworkChuck's switch from OpenClaw to Hermes and in Alex Finn's hundred hours of Hermes lessons. Zubair's Jarvis and Hermes are solving overlapping problems from opposite directions: Hermes is an agent harness you give a body to, and Jarvis is a body somebody built around a model.

Interrupting it, and why that is harder than it looks (5:14)

He cuts the assistant off mid number:

"Okay, that's good now, Jarvis."

It stops. Then he explains the rule:

"See, I can always interrupt him by just saying stop or thank you, Jarvis, or whatever it may be."

Or whatever it may be. That is the important half of the sentence. This is not a keyword match on the literal word stop; it is intent recognition on arbitrary phrasing, evaluated while the assistant is still speaking. Which means the microphone stays open during playback, the pipeline handles hearing its own voice without treating it as user input, and there is a fast path that can classify an incoming utterance as a stop intent and kill generation before the slower main path finishes its turn. Barge in is one of those features that sounds trivial and is a genuine piece of engineering.

He closes the segment on the permission model:

"He sees what you're seeing, with permission, of course. Because of course I want to make sure that he sees only the screens that I'm sharing with them."

Per tab, not per desktop. He shared one browser tab, not the whole machine.

The antenna: a second, faster conversational path (5:40)

The fourth control is a mode switch, not a capability:

"If I click on this antenna, this basically switches the model to a live conversation with the GPT real time, which is a low latency model as well."

He clicks it and moves on without demonstrating it, so take it as described rather than shown. But the design intent is clear and worth spelling out, because it explains a tension the rest of the demo has been living with. The default path is a tool using loop: transcribe, think, possibly call tools, possibly search the web, then synthesize speech. That path is powerful and it is slow enough to need a visible thinking state. A speech to speech realtime model, of the kind OpenAI's Realtime API provides, collapses the transcribe and synthesize steps into the model itself and answers in a few hundred milliseconds, at the cost of the deep tool work.

So he built both and put a button between them. Conversation on the antenna, work on the default path. That is a more sophisticated answer than picking one, and it is one of the few architectural decisions in the video you could copy directly.

The bell: the morning briefing (5:52)

The fifth control:

"If I click on this kind of bell button, this is kind of a briefing. He greets me each morning and gives me all of my to-do list and everything else that needs my attention. He has access to my calendar, my email, and everything else, again, like a true assistant."

Described, not shown. He clicks the button and narrates what it does rather than letting it do it. So the calendar integration, the mail integration, and the to do list aggregation are claims in this video, not demonstrations. They are entirely plausible claims, and they are also the three connectors that take the most work to build and the most permissions to grant, which makes their absence from the demo the most notable gap in it.

Swapping the brain by voice (6:16)

Then comes the trick that most separates this build from a wrapper around one API.

"Right now he's operating with Opus 4.8. I can just tell him just through voice to change his brain to some other AI model, right? So, something cheaper. He also has access to my OpenRouter API key that I've given him. So that means that he can access any model in the market, not only the Claude models. So for instance, I could just tell him to switch to Grok, for instance, or Gemini."

And then he does it, live, twice.

"Hey Jarvis, can you switch your brain to Gemini Pro?"

"The cognitive recalibration is complete."

The panel updates to show Gemini 2.5 Pro. He swaps back:

"Hey Jarvis, switch your brain back to... actually, switch your brain to Opus 4.8."

"Claude Opus 4.8 at your service, sir. The finest artificial intellect money could apparently be persuaded to install in a butler."

Then, almost as an afterthought: "We can give him access to Fable 5 or whatever it may be."

There is one thing to be careful about here, because the video's title invites a misreading. The title credits Claude Fable 5, and Fable 5 is what he says he built the system with. The brain answering questions throughout the demo is Claude Opus 4.8, which the panel confirms and which he states out loud. Both things are true at once and they are different claims: Fable 5 wrote the thing, Opus 4.8 ran it during the recording, and Fable 5 is listed as a model he could point it at. If you came to this video to hear what Fable 5 sounds like as an assistant, that is not what you are hearing.

Why build your own assistant at all (7:20)

He steps back for the argument, and it runs in five moves.

First, the market claim: an assistant like this "is something that's going to be a huge need for people if you're interacting with AI tools and you're trying to save time and become more efficient."

Second, the diagnosis, delivered flatly:

"Obviously, 99% of people are extremely unorganized. If they are organized, all of the organization lives inside files and folders inside their computer that are kind of scattered all over the place."

Third, the fix. An assistant that can interact with your AI second brain, or with "an operating system inside all of the files inside your computer," turns scattered silos into something addressable: "all of the knowledge that are living in different silos is all accessible via voice right in front of you where you can interact with it."

Fourth, the trend line. The need grows as your AI usage grows, "as your interaction with AI expands more and more."

Fifth, why voice specifically: "you can be a lot faster, you're not doing anything manually."

It is a coherent argument and it is worth noticing what it is not. It is not an argument for a smarter model. It is an argument that the bottleneck is the interface between you and the information you already have. That framing is what makes the 3D file visualization the centerpiece rather than a gimmick.

Long term memory, and treating it as a time machine (8:29)

"This is something that remembers you cuz it has long term memory. So everything you teach it, or your preferences, the facts, the way you work, every personalization, it lives inside your long term memory, and therefore you can treat it as a time machine. Ask it, hey, what did I do last Tuesday, I was interacting with a client or working on some kind of a lead, please remind me of that."

That paragraph quietly describes two different storage problems bolted together under one name. Personalization memory is a small, slowly changing set of facts and preferences that belongs in every prompt. Episodic recall, "what did I do last Tuesday," is a growing time indexed log that has to be queried, not injected. They have different write paths, different retrieval strategies, and different failure modes. He calls both of them long term memory, and neither is demonstrated.

Hands (8:52)

"It has hands, right? It has the ability to agentic actions. It can capture, it can look exactly at your screen. It can research online. It can basically look for you inside your email and calendar. You can give it so much more access."

And the roadmap sentiment: "these are just like the initial versions, and as we improve more and more, this is going to become even more and more powerful."

Here is the honest ledger of the whole pitch, sorted by whether the camera backed it up.

Seven of eleven shown, four claimed. That is a better ratio than most tool demos manage, and the four that are missing happen to be the four that need the most plumbing.

How he built it, and what it cost him (9:17)

The chapter titled "How I Built It (Claude Fable 5 + Claude Code)" runs about twenty eight seconds. Here it is, essentially in full:

"As far as building something like this, obviously this version that I showed you guys, this particular version, this took me millions of tokens to build and a lot of many, many hours of going back and forth and kind of problem solving. Yours is going to look different obviously, just because your files are different, your business is different, whatever it may be. This is particularly tailored to mine, but I can give you guys the prompts so that way you can get started and build something similar like this for your own."

That is the entire build section. There is no terminal, no repository, no file tree, no prompt text on screen, no architecture slide, and no dependency list. The cost signals are two: millions of tokens, and many hours of iterative problem solving. No dollar figure, no session count, no timeline in days or weeks.

So the named stack, across the whole twelve minutes, is six items long.

And the only interface specification the video actually hands you is the set of things he said out loud. That is not nothing, because for a voice first system the utterances are the API surface. Here is every command spoken to Jarvis on camera, verbatim, in order:

Jarvis                                         wake word, after the ear button

"pull up my Skool community free Skool         vault retrieval, with the
 community email funnel information"           stumble left in, and it worked

"do a search online and find an alternative    live web research, returns a
 to Kit that I'm using right now, which is     spoken answer plus a card
 costing me a lot of money"

"tell me exactly what you're seeing on my      screen reading, per tab grant
 screen right now that I've shared with you"

"okay, that's good now, Jarvis"                barge in, mid sentence
"stop" / "thank you, Jarvis"                   also accepted, per his own note
                                               that any phrasing works

"switch your brain to Gemini Pro"              model hot swap through OpenRouter
"switch your brain to Opus 4.8"                and back again

"read it out loud"                             offered by Jarvis after a
                                               retrieval, never used on camera

What a rebuild has to supply, which the video never names

The next part is this page's reckoning, not the video's. The video names six components. A working version of what is on screen needs at least a dozen, and every one of the following is visibly present in the demo and completely unnamed in it. If you take the free prompt pack and start building, this is the gap you are walking into.

• Wake word detection. Something is running locally, all the time, listening for one word without shipping audio anywhere. That is its own class of tool, and open options exist (openWakeWord, Picovoice Porcupine).
• Speech to text. Fast, streaming, and accurate enough to survive him restarting a sentence mid phrase. Whisper and Deepgram are the usual answers.
• Text to speech. The single most distinctive thing about the demo is the voice, and the video never says a word about where it comes from. Any rebuild has to pick a voice engine and, if it wants this particular effect, a British butler voice.
• Barge in handling. Keeping the microphone live during playback, ignoring the system's own audio, and classifying a stop intent fast enough to kill generation mid sentence.
• A vault index. Whatever lets "the email funnel and all of the connections" come back as one result. That is a file walker plus a link graph plus retrieval, over his actual folders.
• The 3D file visualization. A whole separate front end project, the "3D Brain" half of the product name.
• The card interface. A panel that renders results as durable dismissible artifacts and can promote one into a note.
• A memory store. Two of them really, one for preferences and one for time indexed episodes, per the distinction his own description implies.
• Screen capture with a per tab grant. Plus the plumbing to hand a frame to a vision capable model on request rather than continuously.
• Calendar and mail connectors. The Google Calendar API and Gmail API or their equivalents, with the OAuth work that implies.
• A model router. OpenRouter covers the provider fan out, but something local still has to own the current selection, expose it as a callable tool, and carry the system prompt across a swap.
• An orchestration loop. Tool definitions, tool selection, the thinking state, and error handling when a tool call fails. This is the part Claude Code is presumably good at generating, and it is also the part that most plausibly consumed the millions of tokens.

The free prompt pack, and where the line is (9:45)

The longest single segment after the demos is the offer, and he walks the exact click path on screen:

"This is my free community. I want to put the link in the description. Go ahead and click on that and you can click Join. Again, it's completely for free, so you're not being charged here. You can go to the Classroom section, click on YouTube Resources, and then right here it says Jarvis AI Assistant Prompts inside Claude Code. So this is the prompt pack that I'm giving you guys for free, where now you can go ahead and read the instructions. You can use these prompts as a guide for you to build something for yourself, for your own business."

Then he draws the line himself, and he draws it clearly enough that it deserves quoting in full rather than paraphrasing:

"Now, of course, like I mentioned, the Jarvis version that you're seeing right now here, it's not like you're going to be building this exact one. But if you want to have access to something like this, like a white label version, this is inside my paid community, right? Because this is hours and hours of research and a lot of tokens. So if you want to have access to this and quickly install this inside your computer via zip file, that's inside my paid community. So again, I want to clarify, because I don't want to deceive anybody. The free version inside my free community, these are the prompt packs that you can build something similar to this, but if you want to have access to something like this, right, like a completed version, that's where you can find that inside my paid community, which is called AI Workshop."

The paid product, in his words: "You join inside the Skool community, you see here V3 3D Brain with Jarvis, where you can just download these brain Jarvis skills and you'll be able to basically upload and point it to your own folders, and everything will be up and running for you with the Jarvis on the side here, where you can do basically everything that I showed you here." Plus a roadmap: "there's multiple versions, and version four is coming up next, where I'll be adding more functionalities to this, and again, that's inside the paid community." Plus two adjacent products: a Claude Code masterclass "if you're new to Claude," and a course on starting an AI agency.

One word in there is a real technical signal and not marketing. He calls the deliverable "brain Jarvis skills." Skills is Claude Code's own term for packaged instruction sets an agent loads, which suggests the thing is not a monolithic application but a bundle of agent skills plus the visualization, installed by dropping a zip and pointing it at your folders. That is consistent with everything else in the demo, and it explains how a model swap can preserve the persona: the persona is in the skill, not the binary.

What should Jarvis learn next (11:40)

The last chapter title reads like a roadmap and is actually a question to the audience. His closing:

"Hopefully you guys found this helpful. Again, the free version is inside the free community. Check that out. Maybe get started, build something for yourself. This would be a really fun project. And again, as you interact with it, as you go back and forth with it more and more, you can keep improving it. You can keep adding more functionalities to it to make it your true AI assistant. Anyways, let me know if you guys have any questions in the comments below."

That is the frame he ends on, and it is the honest one for this kind of system: it is not a product you finish, it is a thing you keep extending toward your own working habits. Which is also, conveniently, the argument for why the paid version ships as V3 with a V4 already promised.

Key takeaways

• The assistant is a voice layer bolted to an existing 3D file system visualization. The product name, "V3 3D Brain with Jarvis," tells you they are two components, and reproducing the demo means reproducing both.
• Seven of the eleven advertised capabilities are demonstrated on camera. Four are described only: the realtime voice mode, the morning briefing, calendar and mail access, and long term memory. Those four are also the four with the heaviest plumbing.
• The strongest technical result is the screen read at 4:57, which pulled a site name, repository, language, license, verbatim tagline, global rank, three counts and a weekly delta off a chart dashboard, correctly attributed.
• The most reusable idea is the model hot swap. Model selection is exposed as a tool the assistant can call on its own configuration, spoken in plain language, backed by an OpenRouter key, and the persona survived a swap to another vendor's model. The character is a prompt, not a property of Claude.
• The persona is generated per turn, not canned. Two runs of the same introduction request produced different jokes with the same setup.
• Barge in works on arbitrary phrasing, not a keyword. He cut it off mid number with "okay, that's good now, Jarvis," which implies an open microphone during playback and a fast stop intent classifier.
• Screen access is a persistent per tab grant with on demand capture, not a continuous stream. The confirmation phrase is precise about it: "ask me to look anytime."
• Live web research answered the same question two different ways in one video. Expect that. Neither answer was verified on camera.
• The title credits Claude Fable 5, which is the model he built the system with. The brain answering during the demo is Claude Opus 4.8. Both statements are his and both are accurate.
• The build chapter is twenty eight seconds long and contains no code, no commands, and no configuration. The reproducible artifact is a free prompt pack behind a free community signup, and he says plainly that it will not get you the version on screen.
• Cost, as stated: millions of tokens and many hours of back and forth. He describes the tokens as wasted, not spent.

Chapters

• 0:00 Meet JARVIS, the Live Demo

• 0:29 JARVIS Demo: Talking to My Second Brain

• 3:23 JARVIS Demo: Live Internet Research by Voice

• 4:32 JARVIS Demo: Screen Sharing + AI Vision

• 5:38 JARVIS Demo: Morning Briefing & Brain Hot-Swap

• 7:20 Why Build Your Own AI Assistant?

• 9:17 How I Built It (Claude Fable 5 + Claude Code)

• 9:45 Get the FREE JARVIS Prompt Pack

• 11:40 What Should JARVIS Learn Next?

Notable quotes

"Ladies and gentlemen, allow me to introduce myself. I am Jarvis, Sir Zubair's second brain, chief note wrangler, and the only butler he trusts to remember what he forgot he was building." Jarvis, 0:05

"I am, as ever, disturbingly alive, sir. Point me at the first item on the ledger and let us dispense with the pleasantries." Jarvis, 0:13

"After hours of going back and forth and wasting millions of tokens, this is at a place where it's truly an AI assistant." Zubair Trabzada, 0:33

"Hands free engaged, sir. Say Jarvis and I shall answer." Jarvis, 1:35

"He makes the tools and the videos. I make sure he doesn't lose his own train of thought on camera, which is a full time position." Jarvis, 2:00

"Right then, sir. Here are the cheaper contenders. EmailOctopus claims to run up to 50% cheaper than Kit, and Brevo offers the lowest entry pricing because it charges per email sent rather than per subscriber." Jarvis, 3:52

"A starhistory.com dashboard for Nous Research Hermes Agent, sir. A Python MIT repo. Tagline: the agent that grows with you. Sitting at 21 global rank with 208.4k stars, 37.9k forks, and 1.8k contributors." Jarvis, 4:57

"He sees what you're seeing, with permission, of course." Zubair Trabzada, 5:20

"The cognitive recalibration is complete." Jarvis, on being switched to Gemini 2.5 Pro, 6:46

"Claude Opus 4.8 at your service, sir. The finest artificial intellect money could apparently be persuaded to install in a butler." Jarvis, on being switched back, 7:03

"Obviously, 99% of people are extremely unorganized. If they are organized, all of the organization lives inside files and folders inside their computer that are kind of scattered all over the place." Zubair Trabzada, 7:33

"You can treat it as a time machine." Zubair Trabzada, on long term memory, 8:29

"This took me millions of tokens to build and a lot of many, many hours of going back and forth and kind of problem solving." Zubair Trabzada, 9:21

"I want to clarify, because I don't want to deceive anybody." Zubair Trabzada, 10:45

Resources mentioned

The creator and the video

• Zubair Trabzada, AI Workshop on YouTube, the channel this comes from
• Skool, the platform hosting both his free community (where the Jarvis prompt pack lives, linked from the video description) and his paid AI Workshop community (where the V3 3D Brain with Jarvis install lives)

Named in the build

• Claude Code, the environment credited in the chapter title as where Jarvis was built
• Claude Fable 5, the model he says he built it with, from Anthropic
• Claude Opus 4.8, the default brain running during the demo, also from Anthropic
• OpenRouter, his API key for it is what lets the assistant reach any model on the market
• Gemini 2.5 Pro, the swap target demonstrated live, from Google DeepMind
• Grok, named as another available swap target but not demonstrated
• OpenAI's Realtime API, the "GPT real time" low latency path behind the antenna button

Things the assistant looked at or talked about

• Kit, the email marketing platform he currently pays for, at $39 a month on the Creator plan
• MailerLite, named in the cold open at $7 a month as the cheapest full featured alternative
• EmailOctopus, named in the live run as claiming up to 50% cheaper than Kit
• Brevo, named in the live run for charging per email sent rather than per subscriber
• star-history.com, the dashboard on the shared tab that Jarvis read back
• Hermes Agent from Nous Research, the repository on that dashboard. Covered on this site in NetworkChuck's case for switching to Hermes and Alex Finn's hundred hours of Hermes lessons

Context for the concept

• J.A.R.V.I.S., the Iron Man assistant the whole thing is named and voiced after
• Building a Second Brain, the note taking framework the phrase "second brain" comes from

Not named in the video, listed for anyone rebuilding it (this page's additions, not his)

• Wake word detection: openWakeWord, Picovoice Porcupine
• Speech to text: Whisper, Deepgram
• Calendar and mail: the Google Calendar API and Gmail API
• Tool wiring for agents: the Model Context Protocol, the standard way to hand an assistant a tool belt like the one in Figure 1

An honest footnote

Everything above is what the video contains. This is what it is worth knowing about it.

The title and the demo credit two different models. Fable 5 is what he says built the system. Opus 4.8 is what is answering throughout. He states both clearly, so nothing is hidden, but a viewer who came to hear Fable 5 run an assistant is hearing Opus 4.8 do it.

This is a demo, not a tutorial, and the build chapter is twenty eight seconds. No command, no file, no prompt text, no architecture. Everything in the reconstruction above about how the system is put together was read off its behavior, which is why Figure 1 marks its own uncertainty. If you want the buildable version, the path runs through a free community signup for a prompt pack, and by his own account that pack gets you something in the same shape rather than the thing on screen.

Four of the eleven capabilities are claims. The realtime voice mode, the morning briefing, calendar and mail access, and the long term memory are all described rather than demonstrated. They are plausible and they are also the pieces that would take the most work to actually wire up, which makes their absence the thing a careful viewer should notice.

The live research is not reproducible, and the video shows you that without saying so. Two runs of the same question returned three different vendors and two different definitions of cheaper. Nothing was fact checked on camera. That is a property of live web search rather than a defect in the assistant, but it is worth knowing before you make a purchasing decision on a spoken answer.

Nobody verified the screen read. The star history numbers are impressive as a vision result and they were never checked against the page. The right takeaway is that the reading was detailed and internally coherent, not that the figures are correct.

"Millions of tokens" is the whole cost disclosure. No dollar figure, no session count, no calendar time. And his own word for those tokens is wasted, which is a franker description of iterative agent building than most videos in this genre offer.

What is genuinely notable, and easy to lose in the sales frame. Strip out the funnel and three real engineering results are sitting on camera. A vision model reading a dense dashboard with correct label attribution. Barge in that works on whatever phrasing you happen to use, mid sentence, while the assistant is talking. And a spoken sentence that rewrites the assistant's own model configuration across two providers with the persona intact on the other side. Those are shown working, they are not trivial to build, and they are the reason this is worth twelve minutes even if you never click the community link.