Capri-AI: Cyberpunk Agent for Windows 10 Runs Claude Code and Codex

⚡ The Gist in 5 Seconds

• Capri-AI installs on Windows 10 and newer via a regular Capri-Setup.exe and works without a terminal.
• Tasks go to Claude Code, Codex, or Antigravity of your choice — with your own account or API key.
• File and shell access goes through confirmation, the code is open under MIT, but the repo has only 5 commits.

🔍 What Happened

A project called Capri-AI has appeared on GitHub — a native desktop app for Windows 10 and newer that installs via Capri-Setup.exe with a Start menu shortcut. The author calls it a "cyberpunk agent": you write a task in plain language, Capri builds a plan, calls tools, and reports back on the result. The repository lives at this address, the license is MIT, and the history counts 5 commits.

Inside, there are three layers, and that's the most interesting part of the description. An engine router (claude-code · codex · antigravity, plus "custom adapters") decides whose model backend thinks over the task. The tools layer gives the agent files · shell · web — reading and editing files, running commands, network. The third layer is an "approval gate": before acting on your files, Capri shows the plan, and you press "yes" or "no". The README has ready-made scenarios: "clean up Downloads and sort by type", "find a bug in a project and fix it", "gather all PDFs from a folder into one summary", "back up Documents every Monday". The author separately warns about SmartScreen: the app is new and rarely downloaded, so Windows will show a warning.

🧠 Our Take

The bet here isn't on the model but on a division of labor that the README states outright: "you pick the brain and Capri provides the body". In other words, intelligence is rented from Anthropic, OpenAI, or Google, while the project claims the role of the body — the interface, the tools, and the control. Our assessment: this is an honest and, apparently, inevitable niche. CLI agents can already do all the same things, but they live in the terminal, and half the people with files and projects won't go into a terminal.

What this changes for your choice: you're no longer tied to a single vendor in the interface. Comparing Claude Code and Codex on the same task can now be done in one window instead of two terminals. The price for this is your own account or API key for each engine, meaning the costs go to the provider, not to the wrapper's author. Who benefits: model vendors, who get another sales channel, and users who need results without a console. Who loses — in our assessment, the same vendors: the easier it is to switch engines, the faster people compare and leave.

And the main thing that bothers me. shell in the tools list isn't "read a PDF". It's executing commands on your machine by an agent trained on other people's texts and other people's repositories. Confirmation before an action reduces the risk but doesn't remove it: a plan that looks harmless may drag a chain of commands behind it. Five commits and one unknown author are not the track record with which I'd let an agent loose in Documents.

⚠️ What We Don't Know

The source says nothing about user numbers, stars, external audits, or exactly how the "approval gate" works: whether it blocks individual commands or just shows a list. There are no details about the adapters either — what's needed to hook up your own engine. You'll have to verify it yourself: the MIT-licensed sources are openly available, but one author's voice and 5 commits are weak proof of maturity. I wouldn't recommend installing an .exe from a fresh repository with shell access without reading the code.

Prediction: within a quarter, the project will either gain external commits and releases beyond the first installer, or it will stay at five commits. The verifiable marker is the commit history and the Releases section on GitHub.

Disclaimer: a fan project, not officially affiliated with CD PROJEKT RED, Netflix, Studio Trigger, Anthropic, OpenAI, or Google; the information comes from a single source — the repository's README.

Source: GitHub / boardshogunfluctuate