Hover

AI coding agents at work in a notch at the top of your screen.
Hand a task to Kiro, Codex, Cursor, OpenCode or Claude Code by typing or speaking, and watch it get done.

Install · Getting started · Voice · Privacy · Build · Contributing

https://github.com/user-attachments/assets/8c7b498f-4b31-4d88-8861-74a28c711347

Overview

Hover is a desktop app for Windows, Linux and macOS. It lives in a slim black island at the top centre of your main display (on a Mac, around the camera notch). At rest, the island shows the AI quotas you switched on, the agents at work, or a question an agent is waiting on. Hover over it, click it or press Alt+N (Option-N on a Mac) to open the Agent office. There, each task is a bot at a desk in a small 3D office. On a Mac the quotas sit in the menu bar instead, because the notch leaves no room.

The agents run in the background in the folder you choose, up to three at once. Hover never opens a terminal for them. Every session is kept, encrypted, so you can pick any of them up later.

Features

Start a task	Approve a change
Read the answer	Pick up a past session

_{Screenshots use sample tasks and answers.}

Install

Download the latest build from Releases.

Hover then lives in the tray (the menu bar on a Mac). Only one copy runs at a time: starting it again opens the dashboard of the copy that is already running.

Agents

Hover drives the agents you already have. Install and sign in to at least one:

The office shows which agents it found, and how to install or sign in to the rest. On a Mac, each agent's page in Settings has a Set up row that installs what is missing and opens the tool's own sign-in; Hover never sees the credentials.

Getting started

1. Open the office: hover over the notch, click it, or press Alt+N (Option-N on a Mac).
2. Press the circle at the bottom left, pick an agent, and choose a folder.
3. Choose its tool access: Full, Ask first, Ask always or Read only.
4. Type the task and press Enter.

Each agent has its own page in Settings (the office menu → Settings) for its model, effort and tool access. The effort choice shows once the agent has reported its effort levels, after its first run.

[!WARNING] With Full access an agent can change files and run commands in its folder without asking. Use a folder under version control.

Voice

Hold a shortcut, say a task, let go. Hover shows what it heard, the folder, the agent and its access, then starts a new chat after a five-second countdown (Settings → Voice → Start on its own: Off, 3, 5 or 10 s). With a chat open and its reply box open, hold the shortcut with the pointer over the chat to dictate into the reply instead. Voice is off until you switch it on in Settings → Voice.

1. Add projects (Settings → Projects). Pick a folder, and add the other names you might say for it. Each project has its own tool access; a new one starts at Ask first, never Full. When a task names no project, or the name isn't clear, it goes to the default workspace: Hover in your home folder, made the first time it is needed. You can change that folder and its access.

2. The agent. Voice uses the default agent (the one last picked in the new-task circle) with that agent's own model and settings.

3. Choose speech recognition (Settings → Voice):

   • Local (Phonon) runs on your computer and understands English only. Press Download on the Phonon card. Nothing is downloaded before you do.
   • Cloud (Groq) sends the recording to Groq with your own key from console.groq.com and detects the language. Check key tests it.

   Hover never switches between the two on its own.

4. Cleanup (optional). Gemini, OpenAI or any OpenAI-compatible service, with your key and model, fixes punctuation and filler words. If it fails, the original text is used.

5. Talk. Hold Ctrl+Alt+Space (on a Mac, Control-Option-Space, which is fixed; elsewhere you can change it), speak, and let go. Edit the task to stop the countdown, then press Start. Esc cancels. Settings → Voice → Try it shows what would start, without starting anything.

6. On the card. The folder, the agent, its model and the Kiro Web button each open a menu or switch for this task only (the countdown stops, and Start is needed). The folder menu lists the default workspace and your voice projects. Say "use Kiro Web" or "use cloud agent" (English) and the task goes to Kiro in Kiro Web, with those words left out of the task. With Kiro Web on, a repository pick appears.

On a Mac, voice is Apple's speech recognizer on the Mac instead (no Phonon, Groq or cleanup). macOS asks for the Microphone and Speech Recognition the first time Hover listens.

Local speech requirements

The model itself is 164 MB. The rest is the private Python and PyTorch runtime that Phonon's engine needs; Hover keeps it in its own folder and never touches a system Python. Windows also needs the Microsoft Visual C++ Redistributable (x64); the card says so before anything is downloaded. Remove deletes the whole install.

Computer use

Switch it on in Settings → Integrations. Hover then gives every agent Cua Driver, which lets it see and operate other apps. Settings installs it and asks for its two permissions (Accessibility and Screen Recording), which go to CuaDriver, not to Hover.

• It works in the background: input goes to the app the agent names, without moving your pointer or taking focus. Hover's guard turns foreground input into background input and refuses what would get in your way: input to the whole desktop, raising a window, moving windows, the clipboard and quitting apps.
• Under Ask first, Ask always and Read only, its calls are treated like any other tool call: asked about, or refused.
• The desk card's Screen tab shows the agent's apps live (on a Mac this needs Hover's own Screen Recording permission; the tab has an Allow… button).
• It is for macOS only. On Windows and Linux the switch is off, and says "Computer use needs macOS."

Privacy and security

Everything lives in %APPDATA%\Hover on Windows, ~/.local/share/Hover on Linux and ~/Library/Application Support/Hover on a Mac. There is no account, server or analytics.

• Sessions are encrypted with AES-GCM under a key protected by Windows DPAPI, on a Mac by the login Keychain, or on Linux by the Secret Service (else a file only you can read).

• API keys (Groq, cleanup) are sealed in secrets.dat with that same key, never in plain text. If the key isn't available, a key you enter is kept only until Hover quits.

• Network. Hover itself goes online only for what you switch on:

  • the Cursor and Claude Code quotas, once every five minutes each;
  • the Phonon download;
  • Groq, in Cloud mode;
  • the cleanup service.

  The Kiro quota runs kiro-cli /usage on your PC, and the Codex quota reads Codex's own logs.

• Discord status. Only when you switch it on. Hover hands the Discord app on your computer a line like "2 agents working" and the agents' names, over the local connection Discord opens for this. Discord then shows it on your profile, as its Activity Privacy setting allows. Hover sends no task names, folders or answers, and makes no sign-in or web request for it.

• Voice data. Cloud sends the audio to Groq; Local keeps recognition on your computer. Cleanup gets the text, never the audio. The recording is deleted after it is turned into text.

• Checkpoints are copies of the files in your project folder (what its .gitignore leaves out is not copied), kept in a Git store in the data folder (checkpoints), unencrypted like the files themselves. Deleting a session deletes its copies.

• Kiro Web. A task you send there goes to Kiro's cloud: its prompt and replies. The cloud clones the repo from GitHub, never from your computer, and gets none of your local MCP servers. It keeps running after Hover quits, until you stop it or delete it in Kiro Web.

• Agents run as hidden child processes in the folder you choose, and stop with Hover. OpenCode's server listens on 127.0.0.1 only, with a password made for each start. Claude Code runs as one process per conversation, in its folder, with your own Claude Code settings. Each agent stops after 5 or 15 idle minutes (Settings) and starts again when you reply.

• Sandbox. On a Mac and on Linux the agents are started under srt (on by default; off with Settings → Integrations → Sandbox). They can write only to their folders, their own state and temp, can't read keychains, mail or other apps' data, and reach the network only through srt's proxy, to the hosts their tool needs. If srt isn't installed the agents start unsandboxed and Settings says so.

• Agent browser. Hover talks to each agent's browser tool over a Unix socket only you can use, with a token made for each start.

• Other tools you switch on. Computer use hands agents the right to see and operate your apps; GitHub CLI setup and Create pull request run gh and git on your PC and push to the remote you pick.

Platform support

Where a feature is off, its switch says why. CI builds the Mac app and runs its packaged Go backend against stand-in tools; the native Swift UI, permission prompts, and real desktop behavior still need hands-on Mac validation. See docs/MACOS.md.

Build from source

Requires Go: the version go.mod names (an older Go fetches it for you). Windows needs no C compiler. Linux needs one and the EGL headers, because Gio draws through EGL. The optional mise setup in mise.toml selects Go and Node and provides task aliases; it does not install platform build prerequisites.

On Ubuntu or Debian:

sudo apt install build-essential libegl1-mesa-dev libgles2-mesa-dev

The office is drawn with wgpu-native, which is a separate library beside the program: build.ps1 fetches wgpu_native.dll and make wgpu fetches libwgpu_native.so.

# Windows
.\build.ps1 run              # build and run
.\build.ps1 publish          # hoverai.exe and wgpu_native.dll in .\publish
.\build.ps1 installer        # Hover-Setup-<version>.exe in .\dist (needs Inno Setup 6 or 7)

# Linux
make                         # ./hover-linux
make wgpu                    # lib/libwgpu_native.so; run with WGPU_NATIVE_PATH=lib/libwgpu_native.so
sudo make install            # /usr/local (PREFIX=, DESTDIR= as usual); make uninstall
make package                 # .deb and tarball in dist/

The version is the number in the VERSION file.

# macOS (needs Xcode's command line tools: xcode-select --install; Node for the office page)
scripts/build-macos.sh             # dist/macos-osx-arm64/Hover.app, signed ad hoc

The Mac app is Swift (macos/) around the web office (web/office/), on the Go backend cmd/hover-backend; scripts/build-macos.sh builds all three into Hover.app (HOVER_ARCH=x64 for Intel). CI builds and signs it ad hoc and puts it on each release as a disk image for Apple silicon and one for Intel; it runs the packaged backend against stand-in tools (tests/macos/backend-smoke.py) and nothing more, and the app isn't notarized yet. docs/MACOS.md has the permissions it asks for.

Developer documentation is in docs/development: architecture, testing, profiling and Windows notes. AGENTS.md summarises how the app works for people and AI agents changing it.

Contributing

Issues and pull requests are welcome. Read CONTRIBUTING.md first. Every pull request builds the app on Windows, Linux and macOS, runs the Go tests on those systems, and drives the Windows app.

The macOS support builds on Arz's (@Entourage397) macOS v1.0 for Hover 2.x, which was Swift and C#.

License

MIT