OpenWand Brings AI Co-Working to Your Desktop With Hotkeys and Auto-Context

A new open-source desktop application called OpenWand is trying to eliminate the window-switching and copy-pasting that comes with using AI chat tools alongside real work. The app sits as an overlay on your desktop, gathers context automatically from your active window, selected text, or browser, and streams answers from any model provider without requiring you to leave your current application.

The core interaction model is simple. Press a hotkey, pick an action from a picker, and OpenWand builds the prompt with whatever context you have enabled. The model responds in a floating bubble. For rewrite tasks, the flow is three steps: select text, press the rewrite hotkey, review the result, and paste it back in place. The app supports Windows 10+, macOS 13+, and Linux X11, with Wayland support in progress.

Automatic Context Without the Copy-Paste Loop

Most AI chat workflows require manually copying relevant text, switching to the chat window, pasting it in, typing a prompt, and switching back. OpenWand compresses this to a hotkey press and a prompt selection. The app captures your selected text and pulls from enabled context sources, such as browser content or active window information, automatically.

For visual context, Ctrl+Alt+Q opens a screen snip flow. You draw a region on screen, send the crop to a vision model, and read the answer in the overlay. No screenshot tool, no file upload, no context switch.

The rewrite-in-place feature captures selected text, sends it through a rewrite prompt, shows a preview, and pastes the accepted result back into the original field. The app tracks which field was active when it was invoked, so the paste lands in the right place.

Any Model, Any Provider, Local or Cloud

OpenWand supports any OpenAI-compatible server, local models, and major cloud providers. The app includes a list of over 20 free and trial API sources for users who do not have a provider yet. Credentials are stored in the OS keychain, not in plain-text configuration files.

The app also integrates with Codex CLI and Claude Agent directly. Users can run either agent through OpenWand, follow live progress, review tool activity and file changes, and approve or reject modifications. Conversations can resume across sessions, and history can be imported, optionally synced, or exported.

A built-in memory system stores short- and long-term memory locally. Users can review and delete entries. The memory is never sent to a provider unless explicitly included in a prompt.

Privacy Modes and Prompt Injection Checks

OpenWand offers three privacy levels. Off mode sends context as-is. Built-in mode locally detects structured secrets like credentials, tokens, and payment details. Advanced mode adds an optional 2.8 GB local OpenAI Privacy Filter for names, addresses, private URLs, and account information. The advanced filter runs locally and may need warmup time.

The app also checks captured text for prompt injection attempts before sending. If it detects an override attempt, it pauses and lets the user continue or cancel. None of this filtering requires sending data to an external service.

Addon System With MCP Bridge

OpenWand is extensible through addons that run in isolated Python processes. Each addon declares its author and requested access before activation, and re-prompts when access expands. Addons can hook into context gathering, model tools, response handling, actions, hotkeys, and the UI.

The app ships with an MCP bridge addon that acts as an MCP client. Users list MCP servers in a JSON file, and OpenWand exposes their tools to the model as native OpenWand tools. A separate OpenWand Context Server provides five read-only tools, get_selected_text, get_clipboard, get_active_window, read_browser_page, and take_screen snip, that trusted MCP clients like Claude Desktop and Cursor can use to read desktop context without OpenWand running.

Requirements and Setup

OpenWand requires Python 3.12. The minimum setup is Windows 10+, macOS 13+, or Linux X11 with 4 GB RAM and 2 GB free disk. Recommended specs add 8 GB+ RAM, 6 GB+ free disk, and a microphone for voice features. Local AI models may need substantially more resources depending on size.

The first launch provisions the Python environment and installs dependencies. Subsequent launches go straight into the app. All settings, chats, memory, and configuration stay on the local machine. There is no hosted storage layer. The project is available on GitHub at SunnyLich/OpenWand.