The Chronicle.

The skill picker gained two bulk sources beyond pasting a link. Point it at a local folder or file and every SKILL.md inside is discovered, deduplicated and added to your library; or copy skills straight from another agent you have already set up. Discovery runs in a sandbox that follows no symlinks and skips node_modules, in all eleven languages.

Settings has a new Terminal section: type any font family installed on your system and size it from 9 to 24, and every live terminal re-measures and reflows on the spot. Nerd Fonts are not bundled, but name one you have installed and its glyphs render at once; the default stack even falls back to Symbols Nerd Font, so status-bar icons can fill in on their own.

The camera finally moves: scroll to dolly in and out, drag to orbit, double-click bare floor to reset. And the room is now an MCP surface the agent can drive. Ask Claude to change the palette, swap the pet, throw a ball or grow the room, and it runs through validated board tools, live and logged to the activity feed. Vibe coding, pointed at the room itself.

When one agent hands work to another, the briefing, skills and context prompt now reach OpenCode as well. The server was already writing the handoff; it simply was not injected at launch for OpenCode, which takes its opening prompt through a --prompt flag rather than a positional argument. Fixed on POSIX and Windows alike, with a guard so a --prompt you set yourself is never overwritten.

Opening a project on Windows now uses the native directory dialog, with the address bar and a paste-ready Folder field, instead of the old tree that would not accept a pasted path. It falls back to the previous server-side picker on the web build or wherever the native bridge is unavailable. Choose a folder the way every other Windows app lets you.

Pressing Esc to back out of a command, or of /status, /context and usage views, used to lock the terminal instead of cancelling. Esc now passes straight through to the shell, and the focus grab that settled afterwards is tied to the typing lock, so your keystrokes land where you are looking.

Every pet shared one generic body, so the cat read as a dog. Each of the six breeds now carries its own silhouette: the cat's pricked ears, whiskers and upright tail, the corgi's long body and nub, the shiba's curled tail, the poodle's topknot, along with faithful coat patterns and legs that tuck into the body's curve.

The app's own activity and usage figures were counting events you did not cause: a project restored at launch logged as a new agent, and an agent lost to a crash logged as a removal. Both passive cases are now gated out, so what the numbers show is work you actually did.

The app was forcing --permission-mode default onto every launch, so a mode set in settings.json, whether plan, accept-edits or bypass, was quietly ignored. That override is gone. The permission mode you configure is the one the agent starts in.

Even in auto mode, agents kept stopping to ask for approval. The same root cause behind the permission-mode fix left the auto-accept path broken across every CLI; one shared answer algorithm now clears those prompts for Claude Code, Codex, Copilot and the rest, so auto mode runs hands-free the way it should.

Typing into one agent could throw you to another: an attention-follow, a background spawn, a fresh terminal or a halted agent exiting would each steal focus mid-keystroke. A typing lock now holds your place, so the pane you are writing in stays put until you are done.

Opening Edit Room with four agents could drop you back to two, a stale in-memory copy of the roster overwriting the saved one. Loading now unions the two lists, keeps the room host pinned and re-syncs focus, so every agent and the room layout survive the round trip.

On a multi-monitor setup the alert toast could appear on the wrong display. It is now anchored to whichever screen the app window sits on, and it follows along if you drag the window or unplug a monitor mid-session.

Force-quitting could make the next launch look wiped: signed out, recent projects empty. Orphaned server processes were squatting on the app's home port, so each launch drifted to a new one and lost sight of its saved state. The app now reclaims its port at boot, cleans up its own strays and keeps a native backup of the recents list. Nothing was ever deleted; agents live on disk.

The workspace tour now stops at what matters: project tabs, adding an agent, the Command Center, git and the terminal. The tab strip gets a one-time coachmark the first time it appears, and the whiteboard and sticky note explain themselves on first click instead of stretching the tour.

Open several projects and they pin as tabs across the top of the workspace. All of their agents share a single room, each at its own desk, and switching tabs refocuses the terminal, roster, git panel, file tree and camera onto that project without rebuilding the world. One WebGL context, no reloads.

Long sessions were heating laptops up and eventually crashing the room. Screen transitions leaked a WebGL context each time, the activity feed grew without bound, and idle rooms kept rendering at full speed. All three are gone: contexts are disposed, the feed is capped, and an idle room barely draws.

A new toy in the room actions: a rope that hangs at the corner of your screen and sways on real physics. Click anywhere and it snaps out to strike that exact spot, with a hard crack, a burst of stars and a recoil from any robot you catch. Room clicks are claimed by the rope while it is armed, so you cannot select someone by accident mid-swing.

Zooming to a desk now parks the camera dead behind the agent instead of slightly off axis. The full-screen Command Center no longer covers dialogs like Add Project. And the live file tree survives server restarts and reconnects instead of quietly going stale until the next reopen.

macOS runs unmoved apps from a randomized read-only path, a feature called App Translocation, and it was silently breaking hook events and freezing approval prompts on fresh installs. The app now detects the situation, repairs its hook and MCP configs, and offers to move itself into Applications.

A batch of animation honesty fixes: agents no longer drift to sleep while a tool call is still running, a long command now plays a proper terminal scene at the desk, and reading a file no longer shows the waiting-for-you pose. Two Claude agents working side by side also stop leaking into each other's scenes; every hook event resolves to the agent it actually belongs to.

Alert sounds now live on their own bus with a dedicated Notifications slider in the Sound panel, in all eleven languages. Mute everything for a meeting and, if you flip the new toggle next to the mute box, the app can still ping you when an agent finishes or needs an answer.

Claude sessions already resume natively. Every other CLI now gets a recap on restore: the app writes a short, display-safe summary of the last session into the terminal, so Codex, Copilot, Gemini and the rest can pick up where they left off instead of waking with amnesia.

The packaged app was silently blocked from loading fonts off a CDN, so text rendered thin, terminals spaced oddly and italics were faked. All typefaces now ship inside the app, and the desktop build renders pixel-identical to the web.

Drag a file through the room and the pet notices: ears up, a chase, a pounce on the drop point. Afterwards it leaps to the window sill, gazes out for a moment, then trots home. It now visits the window on its own from time to time too.

An empty terminal area now offers a grid of CLI and model tiles: one click spawns an agent, no dialog, and it remembers your last pick. The same launcher sits behind the lightning button and Terminal in the menu bar, because the desktop app finally has real File, Edit, View, Window and Help menus with working shortcuts for New Agent, Open Project and Open Recent.

Cards on the personal to-do board grew a send action: pick an idle agent from the right-click menu and the card lands in its terminal as a prompt, moves itself to Doing and confirms with a toast. And when an agent finishes or asks a question while its terminal is closed, its roster card now wears a red unread badge until you check in.

The closing summary an agent writes at the end of a turn kept vanishing from history: the transcript is flushed to disk asynchronously, so at the moment a turn ended its final message often was not there yet. The stop event now carries that message itself, and history reconciles it, adding it if missing and never duplicating it.

The agent trash shows six cards before it scrolls and loads portraits lazily. Terminal glyphs no longer render oversized after a round trip through Edit Room; the character cell is re-measured on the way back. And agent speech bubbles no longer float over the code view.

Sending a message while the agent's terminal was showing a selection prompt used to type raw text into the menu and confirm the wrong option. The composer now holds your message, mirrors the prompt's options inline with a nudge to answer first, and delivers the text the moment the prompt clears.

Two launch fixes: Open Project no longer flashes back to the landing screen in a loop while the session restores, and a machine that cannot reach the auth service, offline or with a skewed system clock, now opens your workspace with the saved identity instead of locking you out, with an honest note about why.

Waking a sleeping agent sometimes ran a single line and dropped straight back to sleep. The cause was resuming from an empty or corrupt session record; the server now validates a session before offering it, and the client detects an instant exit and retries once with a fresh one.

In the packaged app, non-Claude CLIs such as Codex and Copilot failed to start the termi-board MCP server because its helper script never made it into the bundle; dev builds worked, installs did not. The script now ships, staged like the hook shim, and a Windows path bug that mangled its location is fixed alongside it.

Screens in the room now show only the agent's real work: actual edits typed out live, a reading scene with the true file name, the real command during a shell run, and a plain working spinner when there is nothing to show. The canned fake code is gone, keyboard sounds fire only on real typing, and reopening an agent no longer replays old activity into the room.

Paste a GitHub link, an npx command or a package name into the skill picker and every SKILL.md found inside is scanned, deduplicated and imported into your personal library, available across all projects. The picker is your own collection now; the prebuilt catalog is retired.

With several projects open, the boundaries now hold: the git panel binds to each project's own repository root instead of a parent repo, terminals open in the project folder, and the live file tree watches every project root at once.

Some scores ran into the billions because cached-token reads were being counted as work. Rankings now count real input and output tokens only, and new accounts get a random agent portrait instead of a blank avatar.

Seven field guides written from daily agent use, free and fluff-free: what vibe coding actually means, how coding agents work, the 2026 agent comparison, a beginner's first-week setup, the tools stack, a multi-agent field guide and how to recover a closed Claude Code session. Start at the Learn hub.

Claude Code, Codex CLI, Gemini CLI, Copilot, Aider, opencode and local models, judged from running them side by side in Termi rooms every day: who the big four are, the honest weak spots, and the free stack that actually works. Read the comparison.

The desktop app now routes the password recovery email back through its own local server. Clicking the link opens the reset form inside the app instead of landing on a dead browser page.

The /demo page now runs a fully scripted agent session in real ANSI output. It fires the same hook events a live agent would, so the 3D choreography, file-card animations and kanban board all update through production code. No model is involved.

When an agent asks a structured question with a list of options, those options appear as numbered buttons in the notification card. Tapping one drives the terminal's own selection, so the answer lands correctly and no digits end up in the chat.

While an agent is working, a live card in the Memory tab tracks every tool call and permission ask in real time. Turns that were interrupted by a new message or an app restart are recovered from disk on reopen and labeled honestly.

Model quota notes, time labels and source labels in the Status tab now follow the app's active language. Previously some strings were hardcoded and stayed in Turkish or English regardless of the setting.

Every approval gate that fired during a turn is now part of that turn's record: which tool triggered it, what you decided, and when it resolved. Open any past turn and the full picture is there, including questions the agent asked through the terminal.

Multi-agent sync mode file locks are now written to disk. A server crash or restart no longer silently cleared them, which was letting two synced agents claim the same file and overwrite each other's work.

The Activity tab now reads each CLI's local session logs: Codex session files, Copilot CLI event records, and OpenCode's SQLite database. For Gemini and Grok, which store usage data in formats that are not readable, the tab says so rather than showing a false zero.

Closing the app mid-way through the local AI model download no longer means starting from zero. Reopen it and the download continues from the last byte using an HTTP range request against the same file.

Bash commands, git checkouts and package installs now appear in the file explorer as they happen. A background watcher on the project root catches every kind of change, not just the files the agent writes through its own tools.

Each provisioned agent now has a cryptographic secret written at setup time. Hook requests that cannot produce it are rejected with a timing-safe check, so a rogue process on the same machine cannot inject false tool events into the room or the agent's memory.

Pricing is locked in: a one-time lifetime license. Pay once, bring your agent, and your rooms are yours for good. No subscriptions, no seats.

We tightened the server end to end: hardened response headers, rate limits on the heavy routes, stricter input handling and a clean dependency audit.

File-tree rows are real buttons now, reachable by Tab and triggered with Enter or Space, with proper labels on the images throughout.

Flaky operations retry on their own with backoff, checkpoints re-snapshot if a save misses, and the server sweeps up orphaned worktrees the moment it boots.

Every room and every agent moved to a cel-shaded look with ink outlines, and a fresh sky paints sun, clouds and a moon behind each palette. The rooms feel alive, and they are no longer too dark.

The whole app speaks both languages with a live switcher. Flip between them anytime, no reload of your room required.

Bring a new agent into a project and it reads the last one's memory, open tasks and recent prompts, then carries on. It works even across different CLI tools.

Removing or exiting an agent now prunes it for good, with no ghosts coming back on reload, and the registry writes atomically so it never tears.

New here? A gentle spotlight tour walks you through the lobby, your first agent and every tab of the Command Center. Replay it anytime from the ? button.

Sign in and start climbing from Hello World to root. Token and XP boards rank the most prolific coders across today, this week, this month and all time.

They get up and use the furniture, speed up when the work heats up, celebrate finished tasks, and their Task helpers show up as little bots, each at its own desk with a monitor, desktop or laptop.

The full cast is revealed: six Core agents from day one, six Specialists unlocked by level, and four Elite architects for the long haul. Summon any of them by @mention.

The terminal now resizes cleanly over the fd-3 control channel, and a spawn-sequence race that occasionally dropped new agents before their desk finished building is gone.

From Milano to Moscow, the full path from a single desk to a living MMORPG of AI agents, drawn straight from the design bible.

Agents now run on advanced model systems, tighter changes, and they check their own work before handing it back.

Shadow-git worktrees now remove cleanly when unchanged, and the checkpoint diff view stays stable across agent switches.

Activity ($/token), Commands, Tasks, Checkpoints, Memory and Graph: a full per-agent control room beside the 3D world.

New agents now materialize through a Matrix digit-rain and a part-by-part desk build, gated on real terminal-stream readiness.

Smoother 60fps when a room is focused, lower idle GPU, and fewer dropped clicks thanks to coalesced raycasts.

The server now runs as an always-on background service on :3001, no more stale processes, no manual restarts.

They served their purpose well. Workspace thread sharing remains, of course, but the public feed retires as we focus on rooms.