Gentle Dot is a Mac assistant that reads the screen and clicks

A browser-tab assistant forgets the chat you had yesterday and stops at the edge of the page. Gentle Dot, released this week in early preview, keeps one running chat with a memory file on disk, listens on a global shortcut, and will click through a real app when the owner grants it. The repository has 137 stars.

The project is MIT-licensed, built by Alan Buscaglia under the Gentleman Programming channel. The macOS build is a Tauri 2 desktop app plus a local Node daemon; a Linux build ships as a .deb and a package for Arch; the same web interface runs in a browser against the daemon on 127.0.0.1:4317. There is no hosted service and no account to make.

One chat, kept on disk

The chat is a single running conversation, not a list of threads to manage. When the context grows, Gentle Dot starts a fresh session seeded with a summary and carries on. Memory sits in a global store the author calls Engram, written to ~/.engram on port 7437, with the project name gentle-dot. Setting GENTLE_DOT_ENGRAM=private gives the assistant its own memory separate from any other Engram project. The assistant's own files, including the sessions and the local access key, live under ~/.gentle-dot with the directory mode set to 0700.

Reading the screen and clicking the screen

With the owner's permission, Gentle Dot sees the screen, opens apps, clicks and types. A pink glow shows where it acts. Each control session needs a native approval dialog in the desktop app, and the stop shortcut and risky-action prompts live outside the agent so a stuck agent cannot waive them. Keyboard shortcut ⌥⇧Esc stops everything at once. Risky actions, named as send, pay, delete and submit, require an extra approval each time. Password managers and System Settings are off limits. A Yolo mode skips per-action prompts for up to one hour, and only the owner can turn it on.

Voice sits in the same chat. Speech runs on the Mac by default; an optional NVIDIA Parakeet local model, listed as supporting 25 languages, keeps a voice session offline, or an OpenAI API key points the transcription at the cloud.

Model and connector choices

Gentle Dot uses the owner's own accounts: an OpenAI or Anthropic API key, or the subscription the owner already pays for, with role profiles and a model switch above the input. Connectors to Notion, Linear and Atlassian go through those vendors' official MCP servers; Discord, Slack and Gmail come with a step-by-step guide; and an owner can import an existing MCP configuration from Claude, Cursor or VS Code. Each connector is read-only or read-and-send, and the assistant shows a preview and asks before any action that writes.

What the security section says is not yet done

The project page states what the assistant cannot yet enforce. The shell the agent has is checked on a best-effort basis only. Starting the daemon from a terminal, or using the browser-only setup, skips the file-signature check on connectors.json, which is why token-based connectors are locked in that mode. On Linux, while the desktop keyring is unlocked, any program running as the owner can read the connector secrets, which includes the assistant's own shell.

There are no published release binaries yet: the repository ships the build scripts for a .dmg on Apple Silicon and the two Linux packages, and the install guide covers the macOS "Open Anyway" step for an unsigned app on macOS 15.

Source

Gentleman-Programming/gentle-dot on GitHub and the author's channel site.