Node-based terminal manager for AI coding agents — tmux-backed terminals and parallel agent sessions as draggable nodes on an infinite pan/zoom canvas. macOS, Linux, and a browser Server Edition.
Windows is a real target now — as an unsigned beta, and that word is doing work. The
session-host phase landed, so the whole core experience works there: agent sessions and their
hooks, managed Claude and Codex accounts, native SSH (port and agent forwarding, ProxyCommand,
Pageant, ssh.exe), live links and co-attach, the phone relay, dev tools, run configurations, the
tray with background mode, and the session reaper and memory panel. What it does not have is a
code-signing certificate, so Windows SmartScreen will warn on the installer, and the build cannot
update itself: it checks for new versions and shows a card that links to the download, and you run
the new installer over the old one. Your projects and settings are kept, and on 0.4.1+ your running
sessions survive the update. Signing is the next phase.
A simulator is a node now. iOS simulator and Android emulator screens live on the canvas: click to tap, drag or scroll to swipe, type straight into the device, Home and Lock buttons, and a device picker that boots one. It is a real stream rather than a screenshot — nothing is encoded while the screen sits still, about 19–30 fps while it moves, first frame in roughly 0.4 s. A run node gets a phone button that opens its device inline under the run controls, so a Flutter or React Native run and the screen it produces are one node; you can pop that screen out into a node of its own and dock it back. The overflow menu mirrors DeviceHub's own: appearance, text size, location, status-bar overrides, shake, Face ID, Open URL, push notifications, installs, screenshots — and what cannot be driven from outside is listed greyed out with the reason instead of quietly missing. iOS needs macOS and Xcode (the node draws the device through a small Swift helper it compiles on first use, about three seconds, once per Xcode build); Android virtual devices work on every desktop OS, Windows included, over the emulator's own interface. Desktop only — a browser or relay tab says so rather than showing an empty frame.
You can pause a whole project, including the sessions you cannot see. The project menu pauses every agent in a project, a group frame pauses the ones inside it, and a selection of two or more pauses just those; the tab carries a count while they are down, and Resume all brings them back. The sessions that are not on screen are the point — a project you are not looking at has no terminal mounted, and those are exactly the ones holding a few hundred megabytes each. Nothing is interrupted: a session mid-turn, or sitting on a question, is marked and pauses the moment its turn ends cleanly, and Cancel pending pauses drops those marks. A pause exits the agent CLI only; the session, your shell and your scrollback stay, and resuming continues the conversation where it stopped. It also tells you the truth about every session it touched — paused, pausing when the turn ends, unreachable, not eligible, skipped and why — so a batch against a disconnected host never reports a silent zero. Right after an app restart only a session last seen idle is paused and the rest are skipped rather than guessed at, because at that moment an idle CLI and one mid-turn look identical. SSH projects included.
The Cursor Agent CLI is a builtin agent. cursor-agent now sits beside Claude, Codex, Gemini,
Copilot, opencode, Grok and Antigravity in the new-session menus, and it arrived with nearly the
whole harness rather than a bare launch command: the RUNNING and NEEDS YOU badges, the unread dot
and the completion notification, session names synced both ways with the node title, resume after
an app restart or a reboot, the context meter, the ⌘M transcript view, context links to other
agents, subagent cards, canvas control so it can open and arrange nodes itself, permission modes,
model switching and --after dependencies. /loop, /schedule, /cron and Branch stay
Claude-only. Every leg was measured against a real cursor-agent build, and docs/cursor-agent.md
says which facts came from a run and which from reading its bundle.
Updates stop nagging, and Restart finally says something. This is the part macOS users will
notice. Clicking Restart to update used to look like nothing had happened while macOS staged
the update underneath; now the button shows Preparing… and Restarting… and the canvas is saved
before the app quits. Automatic checks download quietly — no progress card and no "You're up to
date" flash unless you asked — a staged update is no longer re-downloaded every six hours, and a
failed background check no longer puts up an error card. When a release carries notes, the card
shows the first few lines. Errors that do reach you are plain sentences instead of net::ERR_*.
The canvas and the board got a wide pass. The ⌘M chat view puts the keyboard straight in its
composer and gives every code block a Copy button. Kanban cards carry the session's own agent
actions, and a held launch now shows on the board with a ▶ to start it. You can choose per project
which agent actions skip the confirm dialog. There are commands for the zone thirds and quarters;
⌘+middle-drag pans past a locked canvas and the lock has its own command; the tab strip and
sessions sidebar auto-scroll while you drag something across them. Sticky notes and group frames
can take a solid backing, the model & context meter can be hidden, rendered Markdown gets a font
family, a size and visible table gridlines — and a running Flutter app can switch between checkouts
without rebuilding. nodeterm://open links from other apps exist behind a flag that is off by
default.
Memory and power got stricter. Minimize nodeterm or send it to the tray and, after two minutes,
its terminals release their GPU contexts and parked emulators — then come back warm when you
return. A window merely covered by another app stays warm; that distinction cost a bug on macOS
and is now explicit. On Windows, session-host scrollback is capped at 10,000 lines, which was the
single biggest memory win there. Keep-awake, which holds off idle sleep while a local agent works,
now lets go below 5% battery and re-holds when you plug in, and on a Mac a critical thermal state
releases it too — latched until things cool down so it cannot flap. An unreadable battery level
changes nothing, so a broken reader can never switch the feature off silently.