Jarvis is a desktop AI assistant for Windows, macOS, and Linux that combines voice conversations, text chat, dictation, memory, and tools. It connects to a model server running your chosen AI models, giving you an assistant you can talk to without requiring a cloud AI subscription.

And yes, the name brings Tony Stark's JARVIS to mind. Who wouldn't want a computer assistant you can bounce ideas off, ask to remember things, and put to work? This is an independent project, but that desktop-sidekick appeal is easy to understand. Building the flying suit is still your problem but maybe this will help.

What Makes Jarvis Cool?

Jarvis goes beyond putting a chat box on your desktop. While voice listening is active, it maintains a short, temporary rolling transcript of nearby speech. You can discuss an idea, address Jarvis by name, and ask what it thinks without repeating the entire conversation as a carefully constructed prompt or typing anything.

There is also a text chat window for times when talking to your computer would earn you strange looks, plus dictation for getting spoken words into other applications.

Jarvis keeps a local conversation diary and a structured collection of remembered information, which you can inspect through its Memory Viewer. That gives it a way to carry preferences and useful context between sessions instead of treating every conversation as your first introduction.

Built-in tools cover web searches, weather, time, reading text from screenshots, and local file operations. Optional MCP integrations can connect additional tools, including browser automation and smart-home services. MCP is a standard that lets AI applications communicate with external tools. What you can accomplish depends on the integrations you configure and how well your selected model uses them. Its a voice assitant by nature not an LLM, so you are dependent on that.

Why Would You Want It?

If you already experiment with local AI, Jarvis gives you another way to use those models throughout the day. You can talk through a project, organize rough ideas, ask follow-up questions, or return to something discussed earlier.

Dictation is particularly practical. Rather than typing everything, you can hold a shortcut, speak, and release it to paste the transcription into an application. That can be useful for emails, document drafts, or capturing an idea before it disappears. I found it was handy to use while doing other thing. For example, while typing I asked Jarvis what the weather would be today and it heard me, and responded with it appropriately English gentleman like voice. I never needed to stop typing the project (this article) or look away. So I can see looking forward where this would be a nice productivity booster for me.

Running models locally also gives you more control over where your conversations go. However, local AI does not mean every feature works without a network connection. Web searches, weather, downloads, updates, and connected services can still communicate online. If you configure a remote model server, your requests travel to that server.

You Need Ollama, LM Studio, or Another Compatible Server

Jarvis supplies the assistant interface, voice features, memory, and tools. It needs a separate model server to generate its AI responses.

You can use Ollama, LM Studio, or another server with an OpenAI-compatible API. "OpenAI-compatible" describes how the applications communicate. It does not mean you need an OpenAI account or paid subscription.

For a local setup, install your preferred model runner and keep its server available while using Jarvis. If you use LM Studio, enable its local server so Jarvis can connect. The Jarvis setup wizard guides you through choosing the provider and models.

Installing Jarvis and Choosing Your Models

Download the appropriate package for your system. On Windows, extract the Windows x64 download and launch Jarvis.exe to begin installation and setup.

During our Windows installation, an optional NVIDIA CUDA component added approximately 1.1 GB to the download. CUDA lets compatible NVIDIA graphics cards handle computing tasks beyond rendering graphics. In Jarvis, it can accelerate speech recognition. It is optional and will not provide that acceleration on an AMD or Intel GPU. I'd recommend it if you have the drive space. Otherwise this thing really hits your CPU hard.

The setup wizard checks Ollama and the models Jarvis is currently configured to use. One potentially confusing detail: the first status screen can report models as missing even when Ollama already has other models installed. That does not necessarily mean Ollama is broken. Continue to the next screen, where you can select alternatives. You should also have Ollam running before installing.

Jarvis provides separate choices for the main conversation model and a fast model used for tasks such as interpreting voice requests and selecting tools. During our setup, we selected Qwen 3.5 0.8B for both roles. The wizard also listed larger options and displayed estimated graphics-memory requirements.

Those requirements refer to VRAM, your graphics card's memory, rather than ordinary system RAM. Jarvis also needs resources for speech recognition, so a model running successfully by itself in Ollama does not guarantee that every part of the assistant will fit comfortably in GPU memory together.

Model downloads are separate from the optional CUDA download and can add considerably to the total download size and storage requirements.

What Does “Location Database Not Installed” Mean?

You may also see a warning that the location database is not installed. This refers to the IP-location database used to estimate your location for features such as local weather and time.

It does not mean you need to install MySQL or another database server. Jarvis handles its local conversation storage separately. If you do not need location features, you can leave them disabled.

Using Jarvis for the First Time

Once setup finishes, start with a simple typed message to confirm the model connection works. Open Chat through the tray menu or by right-clicking the desktop face, then try something straightforward before moving on to more complicated requests.

For voice interaction, allow microphone access and make sure the correct recording device is selected. Once Jarvis reports that it is listening, try something like, "Jarvis, help me plan my afternoon." Add details and follow-up questions to see how well your chosen model handles the conversation.

For Windows dictation, the documented default shortcut is Ctrl + Win. Hold the shortcut while speaking, then release it to paste the transcription. Settings includes options for dictation, speech recognition, models, and a Low Power Mode.

If you receive an LLM HTTP 404 error, check the selected model names and provider settings. During our setup, Jarvis reported a missing Qwen 3.8 27B model even though smaller Qwen models were installed. Selecting an installed model in the wizard addressed the configuration mismatch we identified. Finish the wizard to save your choices, then fully exit and restart Jarvis so it reloads them.

A Few Things to Keep in Mind

Jarvis can work with local files, including modifying and deleting them. Start with disposable documents before involving anything important, and treat optional integrations as additional software you need to trust.

The developer primarily works on macOS and notes that Windows and Linux behavior may differ. Performance and reliability also depend on your models, hardware, and the task you ask it to perform. I had a few crashes out of the box and had to tweak the modles, with a bit of trial and error, under settings to get thing stable

Another point to think about. Jarvis is free for noncommercial use. Its license requires separate permission for commercial use, so do not assume a free personal download automatically covers business use.

Geek Verdict

Jarvis isn't too hard to set up, but I wouldn't call it plug-and-play either. Getting the right models selected took some sorting out. Once that was settled, though, it became a lot of fun to play with. There is something satisfying about talking to your own desktop assistant and it talking back -- sometimes quite snarky.

If you plan to use Jarvis regularly, I would recommend a beefier machine with enough memory to run larger, more capable models alongside speech recognition. Smaller models give you a way to experiment, but your hardware and model choices will have a big say in the experience. If your current PC isn't up to it, getting the setup you want could mean a significant investment.

For curious geeks who already have suitable hardware, Jarvis is an entertaining project with useful possibilities. Just budget for the setup time as well as the downloads. Tony Stark made this stuff look considerably easier.