atrium documentation: CLI coding agents across your machines

atrium is a native macOS and Windows app for running CLI coding agents alongside terminals, editors, and browsers in a tiling layout that survives crashes and reboots. Work can run locally, on Linux boxes connected over SSH, or on paired computers. Each machine owns its sessions and files. The iPhone and iPad app, available in private beta, lets you follow and continue that work away from your desk.

What it is

• A tiling mosaic of panes — terminals, agent chats, editors, a markdown editor, a Chromium browser, git source control, search, and task cards — arranged in splits and rooms inside projects, which are themselves grouped into switchable workspaces.
• A resumable environment: pane layout, terminal scrollback, working directories, environment, and active agent sessions are journaled on the location that owns them and restored after quits, crashes, and reboots. A background daemon on each location keeps agents and recurring tasks running even when the window is closed.
• An adapter-driven AI host: Claude Code, Codex, Antigravity, Cursor Agent, Kimi Code, OpenCode, Pi, Grok, Hermes, Oh My Pi, Muse Code, and Goose are supported through adapters. Third-party adapters can be built with the declarative SDK v2 manifest.
• A programmable surface: every concept (pane, room, project, workspace, location, task, browser, memory) is reachable from the atrium CLI, so agents can coordinate across machines, drive browser panes, update task cards, delegate work, and—when you grant it—control desktop apps.

What it is not

• Not an Electron app. Not a terminal emulator with "AI features bolted on". Not a language server / IDE — Monaco and LSP are used as components, but atrium does not try to be a full IDE.
• Linux boxes run the background daemon and tools; the desktop client runs on macOS and Windows.
• Not a reimplementation of any agent. Whether you run one in a terminal pane or an agent chat pane, it is the vendor's real CLI — a PTY in the first case, the tool's own protocol in the second.

How this documentation is organized

• Getting started — install, first run, launching an agent.
• Core concepts — the mental model: workspaces, projects, locations, rooms, panes, agents, tasks, memory, persistence.
• Panes — one page per pane type, with full capability list.
• Agents — launching, the skills & agents registry, +skill / ++agent sigils, activity, messaging, task dispatch.
• Customization — themes, keybindings, appearance, per-adapter environment.
• Adapters — installing adapters and building your own.
• Reference — keyboard shortcuts, CLI, configuration schema, ~/.atrium/ layout, atrium:// protocol.
• Privacy — what telemetry exists and how to opt out.

If you want to skim, Quickstart is the fastest way in.