Quick Answer: The Capri AI agent is a Windows desktop app from the ChannelCooperStart GitHub project. It sits on top of AI assistants such as Claude Code, Codex, and Antigravity, and it can plan tasks and act on your files and apps. You describe the goal in plain language and approve sensitive actions. It is an unofficial fan project, and macOS and Linux versions are still on the roadmap.
Most AI coding assistants live in a terminal and assume you already know where your files are. Capri takes a different bet. It wraps the assistants you already pay for in a desktop interface, then lets them work across your whole machine instead of one project folder. That sounds useful, and parts of it are. But an agent with access to your desktop deserves a closer look before you hand it a delete command. Here's what the project actually offers, what it leaves unproven, and how to set it up so you stay in charge.
What Capri Actually Is
According to the project README, Capri is "a desktop interface and agent layer" rather than an AI model of its own. You choose the brain, and Capri supplies the body: a window, a tool layer, and an approval step between the model's plans and your computer.
The architecture is simple. Your request goes to an engine router, which forwards it to the assistant you selected. The engine plans the steps and asks Capri to call tools for files, shell commands, or the web. Sensitive actions pass through an approval gate before they run.
The "Edgerunners" branding is cosmetic. The README says the name and neon style reference the AI construct from Cyberpunk: Edgerunners 2, and the project is explicitly unaffiliated with CD PROJEKT RED or Studio Trigger. Treat the theme as a look, not a feature.
Setting Up Capri on Windows
Setup is short, but there are two points where people get stuck. Here's the path the README describes:
- Download the build you want. The project ships a
Capri-Setup.exeinstaller and aCapri-Portable.zip. The installer adds a Start menu shortcut. The portable build runs from any folder with no installation. - Handle the SmartScreen prompt. New executables that few people have downloaded yet can trigger Windows SmartScreen. The README's fix is to click More info, then Run anyway. Only do this if you downloaded the file from the official Releases page, and check the source before you bypass the warning.
- Connect an engine on first launch. Capri asks which assistant to use. You sign in to your account or paste an API key for Claude Code, Codex, Antigravity, or another supported assistant.
Step three is where the real decision happens. Your API key or account determines which provider's terms govern your requests, so read those terms before you connect.
Choosing an Engine: Trade-Offs You Should Know
The README lists three engines as supported and one more as available through a custom adapter. That's a meaningful distinction. Claude Code (from Anthropic) and Codex (from OpenAI) are established coding agents with their own behavior, cost models, and limits. Antigravity is listed too, but the README gives no detail on how its capabilities map onto Capri's tool layer.
Because Capri routes through whichever engine you pick, your results depend on that engine's strengths. A model that excels at refactoring a repository may be a weaker choice for sorting a messy Downloads folder. Test each engine on a low-stakes task before you rely on one for real work.
Switching is one of Capri's selling points, and it's genuinely useful when a provider has an outage or changes pricing. The catch is that prompts and behavior won't transfer one-to-one between engines. Expect to adjust how you phrase tasks.
The Permission Model Is the Real Product
This is the part most reviews skip, and it's where the design matters most. The sample configuration in the README exposes three permission modes:
| Mode | What it means | Best used for |
|---|---|---|
ask | Capri proposes each action and waits for your approval | First runs, file moves, anything irreversible |
auto | Capri acts without per-step confirmation | Repeatable tasks you've already tested in ask mode |
read-only | Capri can read and analyze but not change anything | Summaries, audits, and code review |
The common advice is to switch an agent to automatic mode so it stops interrupting you. For a desktop agent, that advice is backwards at the start. Automatic mode removes the one checkpoint that catches a bad plan, such as a file grouping that breaks a project's relative links or a rule that moves a file you still need. Start in ask mode, watch three or four full runs, and only then consider auto for a narrow, repeatable task.
The config also includes an allowed_folders list. Limit it to the folders the task needs, such as ~/Documents and ~/Projects, rather than your entire user profile. A folder scope turns a bad plan from a disaster into an annoyance.
What You Can Realistically Ask It to Do
The README's examples are straightforward: clean up Downloads, find and fix a bug, summarize a folder of PDFs, scaffold a web app, and schedule a recurring backup. Each one fits the plan, act, and report loop the project describes.
Here's where the examples need caveats. "Group everything by type" sounds simple until you hit files with ambiguous extensions, duplicate names, or files that other programs expect in a specific location. "Summarize every PDF" works best when the PDFs are text-based. Scanned documents depend on the engine's OCR capability, and the README says nothing about that.
Automations deserve the most caution. The README's roadmap lists a "Task scheduler" as not yet done, which leaves the status of the recurring backup example unclear. Test a scheduled task by running it manually first, and confirm it does what you expect before trusting it to run unattended.
Known Gaps and Open Questions
A responsible review has to name what's unfinished. The roadmap shows macOS, Linux, a plugin system, a task scheduler, and a mobile companion as unchecked. Windows and voice control are checked. If you need cross-platform support today, Capri doesn't offer it yet.
The FAQ has an unfinished section on data handling. The README says your requests go to the assistant you connect, under that provider's terms, and then includes a comment marked "TODO: describe exactly what Capri itself stores or sends." Until the maintainers fill that in, don't assume Capri stores nothing locally or sends nothing to its own servers. Read the source, check the release notes, or ask in the issue tracker before you use it with sensitive client files.
The project also describes itself as a fan project and says it's not affiliated with any rights holder. That's a reasonable disclosure, but it means the project's long-term maintenance depends on its individual author or team. Watch the release cadence before you build a workflow around it.
Frequently Asked Questions
What is the Capri AI agent?
Capri is a Windows desktop app that connects AI assistants like Claude Code, Codex, and Antigravity to your files, projects, and apps. You describe a task in plain language, Capri plans and runs the steps through the engine you chose, and it asks for approval before sensitive actions. It is an unofficial, Cyberpunk-themed project that is not affiliated with CD PROJEKT RED or Studio Trigger.
How do I connect Claude Code to Capri on Windows?
Install Capri using the installer or portable build from the GitHub Releases page. On first launch, choose Claude Code when prompted, then sign in to your account or paste your API key. Your key controls which provider's terms apply, so check those terms before connecting.
Is Capri safe to let control my desktop?
It can be, if you configure it carefully. Start in ask mode so you approve every action, limit allowed_folders to the folders a task needs, and keep backups before running large file operations. The README says no tool is risk-free, and the data-handling section is still marked TODO, so review what leaves your machine before using it with sensitive files.
Why does Capri ask for approval before acting?
The approval gate exists so you can catch a bad plan before it runs. Models can misread a folder's purpose or choose a destination that breaks other files. Keeping ask mode on lets you see each step, and it's the safest default for file moves, deletions, and overwrites.
Does Capri work on macOS or Linux yet?
No. The README lists Windows 10 or newer as the only supported platform, and the macOS and Linux versions are unchecked on the roadmap. Check the repository's releases page before you plan around either one.
If you're considering Capri, start with one low-stakes task in a test folder. Ask Capri to summarize a set of documents in read-only mode, then compare its plan to what you'd have done yourself. Pass this to a colleague who keeps asking whether a desktop AI agent is worth the risk, and let them see the permission model before they decide.