Atrium adds a persistent workspace for coding agents

Atrium is a new workspace manager built for people using coding agents such as Claude Code, Codex, and Gemini CLI. According to the project’s Show HN post, it is designed as a “control room” where humans can launch agents, inspect their work, and recover sessions after crashes, reboots, or reopening the app.

The project’s main idea is persistence. Instead of treating an agent session like a one-off prompt, atrium keeps the room, the work, and the coordination state available over time. The author says the goal is to give humans and agents a shared environment rather than forcing everyone to rebuild context after every interruption.

Atrium is also built around worktree isolation. The post says each task can launch into its own git worktree with a command such as task launch ATR-N --worktree, which creates a fresh room and binds an agent to it. That setup is meant to let teams run multiple agents in parallel without file collisions, drifting diffs, or conflicts over locked files.

The app keeps review and oversight inside the same workspace. According to the post, users can review attributed diffs side-by-side, leave anchored comments, resolve threads, and decide what reaches main. A live sidebar shows the state of every agent across workspaces, including whether it is working, waiting, parked, or dismissed.

Atrium also exposes process controls for active terminals. The sidebar shows the running command for each terminal, and users can kill a stuck process, replay the last command, or jump directly to the relevant pane. The app sends notifications when an agent is waiting on a human, rather than notifying on every message it emits.

The project includes tools for pausing and resuming work without losing context. According to the post, dismissing an agent snapshots its pane and lets the session resume exactly where it stopped. A stalled agent can also be nudged back to work with a single click that sends a “please continue” message into its session.

Atrium is not positioned as another AI IDE or an unattended orchestration layer. The author says it is meant to keep humans in the loop while making agent work inspectable, redirectable, and durable. The workspace includes files, browser research, tasks, search, source control, pane management, a markdown editor, and a built-in browser pane.

The app is also heavily CLI-driven. The post says every pane, task, agent, room, and browser can be addressed from the command line, and atrium exports $ATRIUM_CLI_PATH into every pane it owns. Commands support --json, which the author says lets both humans and agents script the same surface.

Atrium runs local-first, according to the post. Workspaces, sessions, and scrollback live on disk in ~/.atrium/, code stays on the user’s machine, and there is no atrium account or model key stored by the company. The project currently supports macOS, with Windows and Linux planned.