A repo called Outpostkrewhirl86/Capri-AI has landed on GitHub. Inside: a README and a license, with the entire history fitting into two commits. The author describes Capri as a local-first agent for Windows with a "cyberbanking" interface and an honest disclaimer: it's a fan project inspired by Cyberpunk: Edgerunners, with no affiliation to CD PROJEKT RED, Studio Trigger, Netflix, or Anthropic. The download link points to a Windows build.
The mechanics are laid out in the README itself. The agent runtime lives in a native desktop shell: an IPC event bus connects the HUD overlay, command palette, and system tray on one side to the runtime on the other. You hand it a task in plain text, and Capri breaks it down into tool calls with a live-streamed plan — at any step you can pause, tweak, or abort. The terminal runs in a sandbox with live output and an audit log, Git operations sit behind confirmation gates, custom tools register via JSON schema, and external servers hook in over MCP. Plus a task queue, so multiple jobs run in parallel.
It looks like the author built what Claude Code and Codex themselves are missing: a body. Both have a brain and a terminal, but no HUD, no tray, no hotkey, no window management, no shared memory across sessions. Capri claims exactly that — a model-agnostic layer where you swap backends like cartridges. Our take: the idea is right, because vendor lock-in is currently the biggest headache for anyone running agents on desktop tasks.
What this changes in practice is a question of the cost of failure, not convenience. An agent with access to your files, clipboard, screen, and terminal is the widest set of permissions you can hand a program. The sandbox and audit log are claimed in the README — but claimed in words, not in an audit report. We'd bet the first users will test exactly that, and file issues if anything goes sideways.
Who wins: people who want to keep their context local and skip paying for yet another subscription — you bring your own backend. Who loses: the chat-interface vendors themselves, who are being stripped of their role as the agent's "face." But with two commits and no release history, Capri is still closer to a pretty demo than a product you'd hand your home folder to.
Not a single measurement exists: how much RAM the tray overlay eats, what actually happens inside the terminal "sandbox," how long-term memory is stored and where it goes. Two commits aren't a development history — they're a point on a graph. The list of supported backends is also an open question: "any other" sounds great in a README and terrible in a bug tracker. Test it on a clean VM with a copy of your data.
Forecast: within a quarter, the repo either gets a proper changelog and a few independent security tests, or it stays a beautiful README with no releases.
Disclaimer: a single source — the repo's README; no independent tests or confirmations exist, the project is a fan work with no official ties to the rights holders of the universe.
Source: github.com · post in Telegram