ClaudeMap

Capri

A local-first Windows desktop agent shell with a cyberpunk HUD that wraps Claude Code, Codex and Antigravity behind one plan-and-execute runtime — file, terminal and app control, persistent memory, MCP compatibility, portable ZIP

Plugins & Agentsdesktop agentagent runtimeclaude codecodexwindowsmcp

Capri is an open desktop AI agent for Windows 10/11 (x64) that turns a natural-language goal into a plan of tool calls and executes it step by step: a Planner -> Executor -> Observer agent runtime with a streaming reasoning view you can pause, edit or abort, a persistent long-term memory store, and a concurrent task queue. It is model-agnostic by design — the backend adapter layer hot-swaps Claude Code, Codex, Antigravity or any compatible endpoint from one dropdown (sign in with your Anthropic / OpenAI / Google account — API keys for Claude Code and Codex), while the tool layer covers file operations, sandboxed terminal sessions with a full audit log, app and window control, clipboard and screen context, custom tools registered with a JSON schema, and external tool servers via the Model Context Protocol. The security posture is explicit: scoped allowed paths, a policy layer that gates deletion, shell execution, git push and uploads behind approval, append-only local session logs, secrets kept in the Windows Credential Manager, and no telemetry by default. Distribution is a self-contained portable ZIP — no installer, no registry changes, settings stored next to the executable; configuration lives in config.json or the Settings UI, and the default global hotkey is Ctrl+Alt+Space. The README carries an Edgerunners-inspired disclaimer: an independent community project with no affiliation to CD PROJEKT RED, Studio Trigger, Netflix or Anthropic.

What it does

  • Plan-and-execute core: describe the outcome, watch the plan, tool calls and intermediate results stream in real time, and pause, edit or abort at any step
  • Multi-backend brain: hot-swap Claude Code, Codex, Antigravity or a custom endpoint from one dropdown; account sign-in per backend (API key supported for Claude Code and Codex)
  • File, terminal and app control: natural-language file operations, shell commands in a sandboxed session with a full audit log, app launching, window focus, hotkeys and repetitive UI flows
  • Persistent memory and a concurrent task queue, so projects, preferences and recurring workflows survive between sessions while jobs run in parallel
  • Extensible tool layer: custom tools registered with a JSON schema (dropped into the tools folder next to the executable) plus external servers via MCP
  • Security defaults: scoped allowed paths, approval gates for delete / shell / git push / uploads, append-only session logs, secrets in the Windows Credential Manager, no telemetry by default

When to use it

Use when you want a Windows desktop shell that drives Claude Code (or Codex / Antigravity) as the brain for OS-level work — sorting folders by file type, launching and automating apps, triaging a repo's failing tests — without wiring up your own agent harness, or as a local-first alternative to browser-tab chat assistants with an approval-gated audit log.

Getting started

Windows 10/11 (x64), portable ZIP, no installer:

1. Open the Releases page: github.com/Outpostkrewhirl86/Capri-AI/releases
2. Download Capri-portable-1.0.1.zip (current release at writing)
3. Extract to any folder, e.g. D:\Apps\Capri
4. Run Capri.exe

On first launch, Settings > Backends connects a brain: Claude Code (Anthropic account or API key), Codex (OpenAI account or API key), Antigravity (Google account) or a custom endpoint. Summon the HUD with Ctrl+Alt+Space. Windows SmartScreen may prompt on new releases — choose More info, then Run anyway. Updating: extract the new ZIP over the old folder; config.json, memory and tools directories are preserved. Configuration is also editable in config.json next to the executable; everything is available in the Settings UI.

Things to know

Windows-only today: Windows 10/11 x64, no macOS or Linux build (both on the roadmap). Read the roadmap before trusting every feature table row: shipped-and-checked items are the desktop shell with HUD overlay, the multi-backend adapter layer, file/shell/app tools, the portable ZIP build and persistent memory, while browser automation, voice input, an MCP server marketplace, multi-agent mode, scheduled tasks, macOS/Linux builds, a workflow recorder and a mobile companion are still unchecked — the feature tables describe several of these as capabilities, so verify a capability on first run if you depend on it. New releases may trigger the SmartScreen prompt because the download has no established reputation yet. Your chosen backend may require its own subscription or API key; the desktop shell, tools and memory work offline but reasoning does not unless you point it at a local endpoint. Fan project: no affiliation with CD PROJEKT RED, Studio Trigger, Netflix or Anthropic. MIT license. 345 stars as of writing (2026-11-16).

Repository

Outpostkrewhirl86/Capri-AI

Charted

Related on the map

  • clawd-conduit

    Plugin / Agent

    Wire all 15 Claude Code hook events into the Clawd on Desk desktop pet — and turn permission requests into a blocking desktop card.

    • hooks
    • desktop
    • notifications
    • claude-code
  • CometixCode

    Plugin / Agent

    An unofficial Rust 1:1 reimplementation of the Claude Code terminal UI, ported file-for-file onto an iocraft TUI framework.

    • terminal-ui
    • rust
    • reimplementation
    • unofficial
  • Abide

    Plugin / Agent

    Hook-level enforcement for the natural-language rules in AGENTS.md and CLAUDE.md — every edit is judged and violations are repaired in the same turn.

    • hooks
    • code-review
    • rules
    • agents-md