01What Is Jarvis Ollama?
Jarvis Ollama is a private, Jarvis-style AI assistant that runs 100% offline on your own Mac using Ollama. It talks, remembers, uses tools and searches your notes without sending anything to a server.
Julian Goldie wanted Jarvis from the films, but private. Client notes, business plans and his daily routine should not go to a server somewhere. So the whole thing runs on Ollama, which runs AI models on your own computer.
The build is small:
- One small project.
- One server file.
- One web page.
- No accounts and no API keys.
Julian runs it on an M4 Mac with 16GB of memory, which he calls a normal machine. He says you can get it running yourself in about two minutes once the models are downloaded.
When you ask Jarvis where it is running, it answers that it is running 100% offline on your Mac. That is the whole point of the build: an assistant that feels like the films but keeps your data at home.
02The Jarvis Dashboard: Neural Core, Reactor and Telemetry
The Jarvis screen is split into clear parts, and each one tells you something useful.
On the left is the neural core. It lists every model you have downloaded with Ollama. In this episode that is Gemma 3 at 4 billion parameters, Qwen 2.5 at 7 billion and GPT OSS at 20 billion.
Each card shows:
- The model's size.
- Whether it can use tools.
- Whether it can see images.
- How much of your memory it fills.
A blue bar means it fits comfortably. An orange bar means it takes most of your machine.
In the middle is the reactor. It is cyan when Jarvis is waiting, amber when he is thinking and red when he is listening. When he talks, the rings move with his voice.
The voice is an open source British voice that runs on your Mac, so it starts instantly. Jarvis speaks each sentence as soon as it is written, so you never wait for the whole reply.
Down on the left is telemetry: CPU, memory, tokens per second and how long the first word took. On Julian's Mac, Qwen runs at about 20 to 30 tokens a second.
03How to Set Up Jarvis With Ollama
Julian walks through the full setup in a few short steps. You need a terminal and a browser, and nothing else.
- Install Ollama from ollama.com.
- In your terminal, pull Qwen 2.5 at 7 billion parameters with ollama pull.
- Pull Gemma 3 at 4 billion parameters the same way.
- Pull the small local embedding model that powers the second brain.
- Go into the Jarvis folder and type node server.js.
- Open your browser at localhost.
- Watch the boot sequence, then the dashboard loads.
That is the whole install. There are no accounts to create and no API keys to paste in.
Once it is running, Julian's first move is to fill the brain folder. Drop in five plain markdown notes about your business:
- Your goals.
- Your clients.
- Your content plan.
- Your team.
- Your weekly routine.
They appear in the galaxy as soon as Jarvis starts. Then ask him about your goals and watch the right note light up.
04The Second Brain Galaxy and Memory
Julian's favourite part is the second brain. The galaxy on the right of the screen is his notes. Every dot is a markdown file, such as business notes, content plans, ideas, his team and his routine.
Each colour is a folder.
Jarvis turns every note into an embedding with a small local model, then plots them in 3D. Notes about similar things sit close together. You can drag the galaxy to spin it and hover over a dot to see the file.
When you ask a question, such as what your Q4 goals are, Jarvis finds the matching note and it lights up in the galaxy.
Memory works the same way. Here is how it goes in the episode:
- Julian tells Jarvis to remember that he records videos at 2pm.
- That fact is saved to a file on his Mac.
- It joins the galaxy as a new red dot.
- Later he asks when he records videos, and Jarvis answers 2pm.
Because the fact lives in a file, it is still there after a restart. Julian's tip is to tell Jarvis three facts about how you work, each starting with remember that. Now he knows you.
05Jarvis Tools: Timers, Notes and Opening Apps
On a model that supports tools, like Qwen 2.5, Jarvis can do things as well as talk. That is why the neural core shows which models can use tools.
In this episode, Julian tries three tools:
- Focus timer: He asks for a 25-minute focus timer. A timer appears at the top and counts down. When it ends, Jarvis tells you out loud.
- Writing notes: He asks for a note called client onboarding checklist with the steps for welcoming a new client. Jarvis creates and saves it as a new markdown file, and a new dot appears in the galaxy.
- Opening apps: Say open calendar, and the Calendar app opens on your Mac.
Every tool call shows up in the transcript on the right in amber. You see the tool name, what it was given and what came back. So you always know exactly what Jarvis did.
To talk instead of type, hold the space bar or click the mic, then let go when you are done. Jarvis answers in his British voice, sentence by sentence.
06The Privacy Bar: Proof It Stays Offline
At the top of the Jarvis screen is a privacy bar. It is there so you do not have to trust a promise. You can see it for yourself.
Here is what it shows:
- A label that says offline.
- A count of every call to a local model.
- Cloud calls, which stay at zero.
- Data sent, which stays at zero bytes.
This is the main reason to build Jarvis on Ollama instead of a cloud chatbot. Your client notes, business plans and daily routine are the kind of things many people would rather not upload. With this setup, the models, the voice, the embeddings and the memory all live on your Mac.
It also means there is no running cost for API calls, because there are no API calls. The title calls it free forever for that reason: once the models are downloaded, everything runs on your own hardware.
If you work with sensitive client information, the privacy bar is a quick way to show yourself, or a client, that nothing left the machine.
07What Mac You Need and Which Model to Pick
The model files tell you how much room you need. In Ollama, Qwen 2.5 at 7 billion parameters is 4.7 gigabytes. Gemma 3 at 4 billion is 3.3 gigabytes.
GPT OSS at 20 billion is about 14 gigabytes.
On Julian's 16GB Mac:
- The 4 billion and 7 billion models feel smooth.
- The 20 billion model runs, but it fills most of the memory, which is why its bar is orange.
- More unified memory lets you load a bigger brain.
Julian keeps Jarvis on Qwen 2.5. His reasons are simple:
- It is quick.
- It can use tools.
- It leaves plenty of memory for everything else on the Mac.
A few tips to avoid trouble:
- Pick a model with tools if you want timers, notes and app opening.
- Watch the memory bar. Orange means you are near your limit.
- Do not skip the embedding model, because the second brain needs it.
08Who Jarvis Ollama Is For
Jarvis Ollama suits anyone who wants a voice assistant that knows their business but keeps everything private.
It is a good fit if you:
- Keep client notes or business plans you do not want on a server.
- Want an assistant that remembers how you work after a restart.
- Want to search your own notes by asking a question out loud.
- Have a normal Mac and do not want to pay for API calls.
- Like seeing exactly which tool was used and what it returned.
In short, it is a private assistant with a second brain, memory, tools and a voice, all on your own Mac.
Julian mentions that the full Jarvis project, the second brain template and the prompts from this episode are inside the AI Profit Boardroom, along with a short guide for each one and the faceless video skill he used to make the video. There is also a community of people building their own versions, so if you get stuck, someone has usually solved it already.