Skip to content

Capri brings a local-first desktop AI agent to Windows

A new MIT-licensed GitHub project wraps Claude Code, Codex and Antigravity in a desktop shell that can touch files, apps, the terminal and the browser.

Draftship Editorial3 min read

A one-day-old repo with a cyberpunk pitch

On 14 November 2026 a repository called Capri-AI appeared on GitHub, describing Capri, a local-first desktop AI agent for Windows with a neon overlay and a look borrowed from Cyberpunk: Edgerunners. The project reports 345 stars and 53 forks on the day it was created, and it is published under the MIT license. The project README frames it as a personal Jarvis: software that sits on the machine, takes a goal in plain language, plans the steps, calls tools and repeats until the job is done.

What it does on the machine

The README groups the capabilities into four areas. On the system side, Capri can search, read, create, move and rename files, run shell commands in what the project calls a sandboxed session with live output and an audit log, launch applications and focus windows, and drive a browser through a controlled context. It can also take the clipboard or a screen region as optional context.

On the agent side, a planner, executor and observer loop breaks a request into tool calls. A streaming view shows the plan and intermediate results, with pause, edit and abort available at each step. There is a task queue for running several jobs at once, a memory store that persists between sessions, and a command palette built around the keyboard.

Swapping the brain

Capri does not ship its own model. It wraps external agent backends behind a single interface and lets the user switch between them from a dropdown. The README names Claude Code, Codex and Antigravity, and says a custom adapter can point Capri at any compatible endpoint. Setup differs by backend: an account sign-in or API key for Claude Code and Codex, a Google account for Antigravity.

Configuration lives in a config.json file next to the executable and in a settings UI. The documented example sets a default hotkey of Ctrl + Alt + Space, a memory retention period of 90 days, a list of actions that require approval (delete, shell, git push, network upload), and a sandbox list of allowed directories. The README states that API keys and tokens go into Windows Credential Manager, that no telemetry is collected by default, and that every action is written to a local append-only log.

What this suggests for people building apps

The repository is explicit that this is not a coding tool only. It advertises delegation of refactors, bug hunts, test generation and repository analysis to whichever coding backend is connected, plus git diff, commit, branch and review through natural language behind a confirmation gate. Custom tools are registered by dropping a JSON file with an input schema into a tools folder, and external tool servers can be connected over the Model Context Protocol.

This suggests the interesting part is the shell rather than the model: a local runtime that holds approvals, path scoping and an audit trail around agents that otherwise run with whatever permissions they are given. Whether the implementation matches that description is something a reader would have to check in the code.

Limits and open questions

Support today is Windows 10 and 11 on x64, distributed as a portable ZIP with no installer and no admin rights required. The README warns that Windows SmartScreen may prompt on new releases. macOS and Linux builds are listed on the roadmap as unfinished, along with scheduled tasks, a multi-agent mode, an MCP server marketplace and a mobile companion.

Two items appear both as shipped features and as unchecked roadmap boxes: browser automation and voice input. The README's own FAQ says the desktop shell, tools and memory work offline, but reasoning needs a connected backend unless Capri is pointed at a local model endpoint. Capri itself is free; the backends may require their own subscription or API key. The star and fork counts are the project's own figures from a single day, and the README does not mention an independent security review. The project states it is not affiliated with CD PROJEKT RED, Studio Trigger, Netflix or Anthropic.

For AI app builders

Capri's model-agnostic design is the same decision anyone adding an AI feature faces: pick a backend and keep the ability to change it. Draftship's guide to the AI chat and model choice covers that side for apps built on Draftship, which has no connection to Capri.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.