Give your Windows desktop a JARVIS-style operator: a hands-on guide to Capri

Capri (ChannelCooperStart/capri-ai, 371 GitHub stars two days after its first commit) is a Windows desktop agent that puts Claude Code, Codex or Antigravity behind a neon, Edgerunners-inspired interface. This guide installs the v1.1.2 build, connects an engine, runs two real desktop tasks through its approval gate, and locks it down to one folder.

• Who: Capri, by ChannelCooperStart on GitHub — a cyberpunk-styled desktop agent and interface layer on top of Claude Code, Codex, Antigravity and other assistants via custom adapters.
• What: a Windows app (installer or portable) that plans tasks, uses file, shell, app and web tools, and asks you to approve sensitive actions before acting.
• When: repository created October 27, 2026; version 1.1.2 released October 29, 2026; 371 GitHub stars and 66 forks, verified October 29, 2026.
• Downloads: Capri-Setup-1.1.2.exe (about 117 MB) or Capri-Portable-1.1.2.zip (about 12.6 MB), from the project's Releases page. Requires Windows 10 or newer.
• Cost and licence: the app itself is free to download; you still need an account or API key for the engine you connect, and GitHub reports no licence for the repository.
• Not yet: macOS and Linux versions, a plugin system and a task scheduler are all on the roadmap, not shipped — so recurring automations are not a feature you can rely on today.

Why this matters now

Claude Code, Codex and their peers are powerful, but they live in a terminal and usually inside one project folder. Capri gives them a body: a real Windows app that works across your whole desktop — files, projects, apps and tasks — while you stay the approver. It appeared on GitHub on October 27, 2026 and picked up 371 stars in two days, which says less about maturity than about appetite: plenty of people want an agent they can talk to instead of type at. This guide tests what the v1.1.2 build actually does, and where it stops.

What you'll need

• A PC running Windows 10 or newer. There is no macOS or Linux build yet.
• An account or API key for one supported engine: Claude Code, Codex or Antigravity. Capri is the interface and agent layer, not the model — your requests are processed by whichever engine you connect, under that provider's terms.
• A throwaway test folder with a few expendable files. Do your first runs there, not in your real Documents folder.

Step 1 — Download the v1.1.2 build

Open the project's Releases page and download the latest release, v1.1.2. There are two formats:

• Installer — Capri-Setup-1.1.2.exe (about 117 MB). A regular installation with a Start menu shortcut.
• Portable — Capri-Portable-1.1.2.zip (about 12.6 MB). Runs from any folder, no installation needed.

If you just want to try it, take the portable build: it leaves nothing installed and is trivial to delete afterwards.

Step 2 — Install it and get past SmartScreen

For the installer, run Capri-Setup.exe, follow the wizard, then open Capri from the Start menu. For the portable build, unzip it into a folder of your choice and run it from there. A new, not-yet-widely-downloaded app can trigger a Windows SmartScreen prompt; the project's own guidance is that if you downloaded it from its Releases page, you click More info → Run anyway. That is a judgement call about trusting a three-day-old project — the portable build in a test folder is the low-commitment way to make it.

Step 3 — Connect your AI engine

On first launch, Capri asks which assistant to use. Sign in to your account or paste your API key for Claude Code, Codex, Antigravity or another supported assistant. This is the one choice that matters: Capri routes your request to the engine you selected, the engine plans the steps, and Capri executes them through its tool layer. Switching engines later is a settings change, not a reinstall — the app, not the model, is what you are evaluating here.

Step 4 — Run your first task: clean up a folder

Use the test folder you made, and give Capri the example task straight from its own README:

Clean up my Downloads folder and group everything by type.

Point it at your test folder instead of the real Downloads folder for the first run. Watch what happens: Capri scans the folder, proposes a plan, and — critically — asks you to approve it before moving anything. Approve, then check the result yourself: are the groups what you expected, and did anything get moved that should not have been? The approval gate is the product here. An agent that moves files without asking is a liability; one that shows its plan first is a tool you can learn to trust in stages.

Step 5 — Run a second task: summarise a set of PDFs

Drop three or four PDFs into the test folder and try the README's second example:

Summarize every PDF in this folder into one document.

Capri reads each file and writes a single combined summary. Read that summary against one of the originals before you trust it: check a number, a name and a conclusion. Desktop agents fail in the same quiet ways chatbots do — a plausible summary with a wrong figure — except now the output lands in your files, where it looks authoritative. The check takes a minute and is the difference between delegating and abdicating.

Step 6 — Lock it down to one folder

Before you point Capri at anything real, restrict it. The project's example configuration shows the shape of it:

# Example config
engine: claude-code        # claude-code | codex | antigravity | custom

permissions:
  mode: ask                # ask before acting | auto | read-only
  allowed_folders:
    - ~/Documents
    - ~/Projects

interface:
  theme: cyberpunk

Keep the mode on ask — the other two modes are auto and read-only, and auto removes the very gate that makes the app safe to learn with. Then narrow allowed_folders to exactly the folders the task needs, and no more. If you only ever want it tidying one project, give it one project. An agent's blast radius is a configuration choice, and this is where you make it.

What you built

A desktop operator: Capri installed (or running portably), connected to the engine of your choice, proven on two tasks in a test folder, and confined to the folders you named with approval required before it acts. That is a usable setup — not a demo, a workflow you can repeat on a real folder when you are ready, one folder at a time.

Honest limitations

• Windows only, for now. macOS and Linux are roadmap items, not downloads. If you are not on Windows, this is a project to watch, not to install.
• Very young. Three days old and 371 stars is momentum, not a track record. Expect rough edges, and do not be surprised if a future release changes behaviour.
• Recurring tasks are not shipped. The README's “every Monday, back up my Documents” example reads like a feature, but the task scheduler sits unchecked on the roadmap. Treat scheduled automations as promised, not present.
• It is a fan project. Capri is an independent, unofficial project inspired by Cyberpunk: Edgerunners, not affiliated with or endorsed by CD PROJEKT RED, Studio Trigger or Netflix. The neon is a theme, not a partnership.
• Your data goes to your engine's provider. The project itself notes its data handling is still a to-do in its own FAQ. Keep sensitive folders out of allowed_folders until that is documented.
• GitHub reports no licence for the repository. If licensing matters to how you use or redistribute it, that gap is worth knowing before you build on it.

The takeaway

Capri's idea is sound and its discipline is right: show the plan, ask first, confine the scope. What it is today is a promising Windows front end for engines you already pay for — best tried portably, in a test folder, with the approval gate on. Give it the one folder it needs, and it earns the next one.