-
name: visioncli
-
description: VisionCLI is an open-source (MIT) AI co-assistant for developers. You hold a key and ask, by voice or text, about the file, error or window on screen. It answers through the AI coding CLI the developer already uses (Claude Code or Gemini/Antigravity). It can also carry out tasks by clicking and typing in other apps, and the user approves every action.
-
snapcraft: visioncli/packaging/snap/snapcraft.yaml at main · thethirdsourcers/visioncli · GitHub
-
upstream-relation: I’m the upstream author (publisher: Third Sourcers, https://thethirdsourcers.com)
-
supported category: IDEs / terminals and developer tools
-
reasoning: Strict confinement can’t work, for these reasons:
- It runs the user’s own developer CLIs from the host (claude, agy, gemini) with their existing installs and sign-ins. It also uses tmux to type into a running CLI session.
- It drives other desktop applications. It lists and captures windows (xdotool, wmctrl, maim, grim) and finds and presses buttons through AT-SPI.
- Hold-to-talk reads key presses from /dev/input (the user joins the input group).
- It captures the whole screen and the window under the mouse, across all apps.
There are no strict interfaces for running arbitrary host programs, which is why it needs classic confinement like other developer tools.
Website: https://visioncli.thethirdsourcers.com
Thanks!