nachosanbenito.com
Jarvis
Get in touch
Jarvis is idle
Madrid--:--

Jarvis

A butler that lives on my Mac. It listens, decides who should do each job, and uses the Mac for me.

See its hands work

The whole tour in 107 seconds.

Start here: the island, the voice, hands-free, dictation and teacher mode, in one short film. It has sound.

The promo film. Recorded on my Mac, with Jarvis's real voice. Controls let you pause or go full screen.

Hold ⌘M and talk.

Hold ⌘M like a walkie-talkie, or double-tap it for hands-free: then saying “Jarvis” is enough, and two minutes of silence turns it off. Speech is recognised on the Mac.

Real recording. A question goes into the HUD and the sphere changes colour while it thinks and answers.
Island beside the notch
Shows what it is doing. Drag it along the edge of the screen; hover it for the full card.
Dictation
Say “write the following” and it types what you say into any app.
Receipts
Every order leaves a short receipt in Obsidian.

Then it decides who does the job.

Everyday orders take a fast lane with no model at all. The rest go to Claude Sonnet, with Gemini as the backup if Claude is down.

Roast my idea
Believer, sceptic, investor and judge give a verdict: build, fix first or kill.
Grill me
One question at a time until the plan has no holes, then it writes the spec.
LLM council
Five advisors debate a decision.
Deep research
A web report with sources, saved to Obsidian.
Prompt builder
Turns a rough task into a reusable prompt.
Content cascade
One note becomes a LinkedIn post, a thread and a newsletter paragraph.
43 buttons in 8 groups. Each one is a skill: a written method Jarvis runs through Claude Code.

It answers out loud.

Replies come back in a British butler voice that starts speaking in about 0.1 s, while the rest of the answer is still being written. One click plans the day: classes, deadlines, events and three priorities, written into today's Obsidian note.

Real recording. Plan my day, pressed in the HUD. Schedule and priorities on screen are example data.

Also one click: close the day, score the week from 1 to 10, sort the inbox into urgent, reply and ignore with drafts (it never sends), find internships that fit, and check a CV against a posting.

Then its hands use my Mac.

Say "open Spotify" and a pointer flies to the Dock and clicks. Jarvis reads any window through macOS Accessibility, so it is not limited to its own app.

Open Spotify. The pointer flies to the Dock icon and clicks it.
Search lofi beats on Spotify. It types the query inside the app.
Message: running late. It writes the text, says what it will send and counts down; a click or “stop” cancels.
Explain it to me
Teacher mode: the pointer glides to each step, types at the pace of the voice and leaves a checklist.
Learns recipes
When a plan works, it saves it and repeats it next time without calling a model.
Scroll, volume, media keys
Play, pause, next and volume by voice. Before acting it says what it will do and counts down on the island, so you can stop it.

Built on Plip (MIT), an open-source Mac agent. Every clip on this page is a real screen recording; the hands clips show an earlier version.

Private project, runs on my Mac.

It holds my calendar, mail access and account keys, so there is no public repository. If you want to see how it is built, ask.

  • ⌘M to talk
  • 0.1 s to the first word
  • ~1 % CPU when idle