habblaud

---

You open Claude Code in several projects, launch subagents, leave tasks running… and lose track of who's doing what. Habblaud turns all of that into an office you can take in at a glance: who's typing, who raised a hand because they need you, who delivered their work and left, who went for a coffee while waiting for the next instruction — and how much of each account you've already used in the 5-hour session and in the week.

And when two or more agents have nothing to do, the office gets a social life: each one has its own personality, and they watch soccer together in the lounge, play video games and ping-pong, gossip in the kitchen, freshen up at the mirror and bet coins on rock-paper-scissors. See Social life.

Everything runs on your machine, reading the files Claude Code itself already writes (plus only the alias lines in your shell, to give each account its letter). Nothing leaves the computer: Habblaud doesn't read credentials or make external calls.

Contents

• What it looks like
• Installation
• How to use
• Accounts and usage (5-hour and weekly)
• Configuration
• Habblaud World
• How it works
• Privacy and security
• Development
• Troubleshooting
• Disclaimer
• License
• Social life

What it looks like

Overview. The building has a lobby with elevators, a kitchen, restrooms and a lounge; each project with an open session gets its own room along the hallway. At the top, the counters and each account's usage; on the left, the rooms and agents; at the bottom, the activity feed.

The lights go out. When you close a project's last session, the last character walks to the light switch, turns off the lights and leaves by elevator — and the room is taken apart, becoming a garden until a new project arrives. If the garden ended up between two rooms, the farthest room moves there: it's set up in the empty spot, everyone walks over to it, and the old address goes dark and is taken apart. That way the building has no gaps and shrinks on its own.

1. The room is set up, still dark	2. Someone turns on the lights and everyone works	3. The last one leaves and turns off the lights	4. The room becomes a garden

The images above use demo mode (made-up projects, people and accounts).

Installation

Requirements

• macOS (tested) or Linux.
• Claude Code installed, with one or more accounts (2.1.287 or newer for the mod).
• Node.js 22.12+ and npm (developed with Node 24; Docker already uses Node 24).
• Docker Desktop (or Docker Engine with Compose v2), if you're going to run it in a container.

1. Get the project

git clone https://github.com/marmottajr/habblaud.git
cd habblaud
npm install

npm install only downloads build tools (Vite, TypeScript, tsx…). The Habblaud server has no runtime dependencies: it only uses Node's built-in modules.

2. Start Habblaud

Option A — Docker (recommended for keeping it always on)

npm run docker:up

The script detects your Claude Code accounts, mounts only the folders it needs (read-only), builds the image and starts the container — which restarts on its own along with Docker. Open http://localhost:4747.

Option B — Node, without Docker

npm run build
npm start

Open http://localhost:4747. For development, use npm run dev (server + Vite with hot reload).

Now open Claude Code in any project and watch your agent arrive by elevator. 🎉

3. Install the Habblaud mod in Claude Code (recommended)

npm run mod:install                        # o mod e os plugins de permissões e de mensagens, em cada conta
npm run mod:install -- --sem-permissoes    # sem responder permissões e perguntas pelo escritório
npm run mod:install -- --sem-mensagens     # sem mandar mensagens aos agentes pelo escritório

Habblaud ships a mod (a plugin that runs inside Claude Code) and installs it in each account, through claude plugin itself, from this folder. With it:

• each account's 5-hour and weekly usage shows up live in Habblaud;
• the terminal shows a line when another session needs you (permission or answer);
• the /habblaud command becomes available in Claude Code;
• you can approve or deny permission requests from the office ("Do you want to…") and answer the agent's questions, with the habblaud-permissoes plugin (see Answering from the office);
• you can send messages to the agent from the office, as if you were typing in its terminal, with the habblaud-mensagens plugin (see Sending messages).

It requires Claude Code 2.1.287 or newer (claude --version). The mod runs inside each session, with your permissions, and only accesses what is listed in mod/README.md; to check without running anything, claude plugin validate mod/habblaud shows the events it handles and the calls it makes. Sessions that are already open load the mod with /reload-plugins (or when reopened). If the account had the old setup (below), mod:install removes the tap and the hook from settings.json (with a backup), because the mod does the same thing. Check with npm run mod:status; to undo: npm run mod:uninstall.

The mod is loaded from this folder. If you move the folder, run npm run mod:install again.

Without mods, live usage and answering from the office come from two installers that edit each account's settings.json (with a backup first):

npm run usage:install    # põe um "tap" na frente do statusline de cada conta, que captura o uso de 5h/semanal
npm run hooks:install    # acrescenta o hook PermissionRequest, para responder permissões pelo escritório

Your statusline keeps looking the same. To undo: npm run usage:uninstall and npm run hooks:uninstall. Both point to this folder: if you move the folder, run them again (until then, the accounts' statusline shows an error). When you update Claude Code, switch to the mod: npm run mod:install removes the old tap and hook (with a backup) so you don't end up with two of them capturing usage or answering the same request.

4. Codex in the office (optional)

If you also use OpenAI's Codex (the codex CLI or the app), its sessions join the same office: each project is a room — the same room as Claude Code, when both are open in the same project — and each session is a character with the CODEX tag. Habblaud finds the Codex folder on its own (~/.codex, CODEX_HOME or ~/.codex*) and reads the conversations, without installing anything. To see Codex live (what it's doing right now, who needs you) and approve commands from the office, install the hooks:

npm run codex:install      # acrescenta os hooks do Habblaud em ~/.codex/hooks.json (backup antes; os outros ficam)
npm run codex:status       # confere

Then open Codex and approve the Habblaud hooks in /hooks: Codex only runs a new hook after you approve it (Habblaud never writes that approval). With Habblaud in Docker, to send messages to Codex sessions, also keep npm run codex:bridge running on the Mac (see Codex).

Open on your phone (optional)

By default Habblaud only accepts connections from the computer it runs on. To open it on your phone (on the same Wi-Fi), with Docker:

printf 'HABBLAUD_BIND=0.0.0.0\n' > .env
npm run docker:up -- --no-build

Then open http://<ip-do-computador>:4747 on your phone (on macOS: ipconfig getifaddr en0; if it comes back empty, ipconfig getifaddr en1). In Node mode, use HABBLAUD_HOST=0.0.0.0 npm start.

⚠️ With this, any device on the network can see the agents' activity (commands, files, session titles). Use it only on trusted networks. The terminal (and with it, answering and sending messages from the office) stays off while the port is exposed. To go back: delete .env and run npm run docker:up -- --no-build.

Phone with login and HTTPS (approving requests)

Instead of exposing the main port, open a second port with a username, a password and HTTPS. Port 4747 stays on the computer only (hooks, plugins and your browser keep working without a login); the remote port goes to the network and every request on it requires a login. What separates the two is the port the connection arrived on, never the Host header. Nothing of this exists until you run the setup: without it, no extra port is published.

npm run remote:setup   # usuário, senha, nome do endereço e (opcional) um endereço próprio
npm run docker:up

remote:setup writes only the password hash (scrypt) to .env, creates a local CA and the server certificate in ~/.habblaud/remote-tls, and leaves a service announcing <nome>.local on the network (mDNS; LaunchAgent on macOS, systemd on Linux). Open https://<nome>.local on your phone (with :4748 if port 443 was taken). The first time, accept the warning, tap Install the Habblaud certificate and trust it (iPhone: Settings › Profile Downloaded › Install, then Settings › General › About › Certificate Trust Settings; Android: Security › Encryption & credentials › Install a certificate › CA certificate), checking the fingerprint the setup showed.

• Restricted CA: it only signs localhost, .local names, your own address and private-network IPs. Even if its key leaked, a certificate for any other site would be rejected by the phone that trusts it. Changing your own address creates a new CA (install the certificate on the phone again). The CA key never goes into the container.
• Own address (e.g. meuclaude.home): goes into the certificate, but the phone only finds it with an entry in the router's DNS pointing to the computer's IP. <nome>.local works without configuring anything.
• What the phone can do: by default, approve and deny permission requests. HABBLAUD_REMOTE_ALLOW in .env accepts permissions, messages, terminal and spawn, separated by commas, or none (view only). The interactive terminal never works from the phone.
• Deliveries (phone, with terminal allowed): the box button in the top bar lists what each agent finished, with a summary and attachments (screenshots, PDFs, files it sent) that open with a tap.
• New agent (the + button): starts a Claude Code session in the background on the computer (claude --bg, in the chosen folder), which shows up in the office and receives messages and permission requests through Habblaud. In a folder Claude Code doesn't trust yet, it opens the Claude app's new-session screen filled in, and you press Enter there. It comes turned off: it needs HABBLAUD_PTY=1 (the same opt-in as the interactive terminal; in Docker, put it in .env and keep npm run codex:bridge running on the computer) and, from the phone, spawn in HABBLAUD_REMOTE_ALLOW too.
• Dictation (in New agent and in the message box): needs whisper-cli and ffmpeg (brew install whisper-cpp ffmpeg) and a multilingual model in ~/.local/share/whisper/ (ggml-small.bin gets more right than ggml-base.bin; the .en ones only understand English). Without them, the Dictate button doesn't show up. The voice never leaves the computer. HABBLAUD_WHISPER_BIN, HABBLAUD_FFMPEG_BIN, HABBLAUD_WHISPER_MODEL and (default ) adjust it.

Without HTTPS the remote port doesn't start; HABBLAUD_REMOTE_INSECURE_HTTP=1 forces plain HTTP (password and cookie in the clear on the network). In Node mode: HABBLAUD_REMOTE_PORT and HABBLAUD_REMOTE_HOST (default 0.0.0.0:4748) and HABBLAUD_REMOTE_TLS_DIR (default ~/.habblaud/remote-tls/server).

Updating

The version in use shows up in the top bar, next to "Connected", and in Settings › About. Every 6 hours Habblaud checks GitHub for a new version (the releases of this repository). When there is one, the green New version badge appears in place of the number, along with a notice and a dot on the settings button. About has the link to what changed and the Check now button (each version's notes are also in CHANGELOG.md). To update:

git pull
npm install
npm run docker:up        # ou: npm run build && npm start (e depois npm run mod:install)

If git pull stops with "Your local changes to the following files would be overwritten by merge: package-lock.json", run git checkout -- package-lock.json and try again: up to 0.4.0, npm install changed that file.

docker:up also updates the mod in the accounts where it's already installed (it never installs it on its own) and tells you: "Mod updated to 0.3.0 on Account D; open sessions: /reload-plugins". In Node mode, run npm run mod:install after updating. To skip checking GitHub, use HABBLAUD_UPDATE_CHECK=0 (in .env, for Docker).

Coming from CodeTown

Habblaud was called CodeTown up to 0.3.2. After git pull, run once:

npm install
npm run mod:install      # troca o marketplace e os plugins codetown pelos habblaud, em cada conta
npm run docker:up        # tira o container codetown e copia os dados do volume antigo para o novo

• Mod: until you run mod:install, the accounts' Claude Code complains about the codetown marketplace (the mod/codetown* folders were renamed). docker:up warns you, but doesn't switch it on its own.
• Docker: the old volume (codetown_codetown-data) is left untouched. After checking that the characters' names came over, delete it with docker volume rm codetown_codetown-data, and the old image with docker image rm codetown:local.
• Local folder: ~/.codetown becomes ~/.habblaud on its own (in the server, in docker:up and in mod:install).
• Browser: preferences and coins move to the new keys the first time the page opens.
• Variables: the CODETOWN_* ones no longer work; rename them to HABBLAUD_* in .env (e.g. HABBLAUD_BIND). The server and docker:up warn you when they find an old one.
• Repository: it's now github.com/marmottajr/habblaud (GitHub redirects the old address). To fix your clone: git remote set-url origin https://github.com/marmottajr/habblaud.git. The local folder can keep its old name.
• settings.json backups made before the switch keep the name settings.json.codetown-backup-<data>.

Uninstalling

npm run mod:uninstall                       # tira o mod, os plugins de permissões e de mensagens e o marketplace
npm run codex:uninstall                     # tira os hooks do Habblaud do Codex (os outros ficam)
npm run docker:down                         # para o container
docker volume rm habblaud_habblaud-data     # apaga os dados do container (nomes, salas, linha do tempo e estatísticas)
docker image rm habblaud:local              # apaga a imagem
rm -rf ~/.habblaud                          # apaga os dados locais (uso capturado, nomes, salas, linha do tempo e estatísticas)

Then just delete the project folder — run mod:uninstall first, or the accounts' Claude Code will start complaining about the missing marketplace. If you used the old way, also run npm run usage:uninstall (restores the original statusline) and npm run hooks:uninstall (removes the permission hook); without them, the accounts' statusline starts showing an error (and the hook, failing silently). mod:install, which removes the old tap and hook, and the old installers and uninstallers leave settings.json.habblaud-backup-<data> copies in each account's folder (e.g. ~/.claude/); delete them if you no longer need them.

How to use

What each character is doing

State	What you see
Working	At the desk, typing. The monitor and a bubble show the activity: 📖 reading, ✏️ editing, 💻 terminal, 🧪 tests, 🌐 searching…
Needs you	Rushes to the desk and raises a hand, with a blinking alert: it's waiting for a permission or an answer in the terminal.
Waiting on shell	Finished its turn but left a command running (tests, build, deploy…), so it stays at the desk with an hourglass flipping over its head and the terminal in progress: it eats popcorn while watching, after 3 min it crosses its arms and spins in its chair, after 10 min it gathers cobwebs and after 25 min it dozes off; when the command finishes, it stands up and celebrates with confetti (or gets a rain cloud, if it failed). After 40 s, if there are coworkers with nothing to do, it may head out to a hangout — with the hourglass over its head.
Idle	Finished its turn and wanders around: coffee in the kitchen, the water cooler, the restroom, the lounge couch, the phone on the beanbag, the mirror. With idle coworkers around, it joins a hangout: TV, video games, ping-pong, a kitchen chat, rock-paper-scissors (see Social life). After 10 min without moving, it dozes off — but a coworker may wake it up for a hangout.
Subagent done	Walks over to the agent that called it, delivers the result and leaves by elevator.
Session ended	Leaves; if it was the last one in the room, it turns off the lights on the way out.

Each character has a unique Brazilian name (Marina, Henrique, Luan…) that stays as long as the session exists, and a colored chip with the account letter (C, D…).

Interface

• Top: counters (rooms, agents, working, subagents, shells running, need you) and a per-account usage card. The shells counter only shows up while some command is running; click it to go to whoever is waiting. Below the logo are the connection and the version in use, which turns into the New version badge when a new version comes out (see Updating).
• Side panel: search, account filter and the list of rooms with their agents and subagents.
• Details drawer: click a character (in the building or in the list) to see activity, tasks, subagents, timeline and stats (tools, tokens, cost, lines changed, model, branch).
• Feed: the latest activity across the whole office.
• Settings (⚙): names, bubbles, how much idle agents wander around, day/night cycle, sounds, browser notifications, demo mode and About (version in use and new version). Help (?): full legend and shortcuts.
• My day (📊): where the day's time went (see My day). Timelapse and Session history (the clocks in the top bar): see Day timelapse and Terminal.

Camera: drag to move, scroll to zoom, double-click a character to follow it.

Shortcuts: / search · F follow the selected one · T terminal · L timelapse · M my day · P next request (permission or question) · O or 0 overview · Esc clear selection · [ side panel · ] feed · arrows/WASD move · + - zoom · ? help.

Office themes

In Settings → Appearance → Office theme, choose:

• Automatic (default): follows the calendar, in local time. Halloween from October 1 to 31, Christmas from November 1 to December 26 and, the rest of the year, the original look.
• None: always the original look.
• Christmas: wreaths, warm lights, trees with presents, Christmas rugs, snow on the windows and Santa hats on the characters.
• Halloween: pumpkins, decorative ghosts, cobwebs, bats, purple lights, themed rugs and costume accessories.

The switch is immediate and the choice is saved in this browser. The themes keep the agents' paths, activities and animations. The day/night cycle is still a separate option.

Languages

The interface is in Portuguese, English, Spanish and French. It follows the browser's language; in Settings › Language you can pin one (the choice is saved in this browser). The texts the server builds (activities, notices, terminal, logs and the install scripts) follow HABBLAUD_LANG and, without it, the system language, except English: many machines and every container come in English by default, even for people who use another language. For the server in English, set HABBLAUD_LANG=en (in Docker, in .env). Any text without a translation shows up in Portuguese.

Day, night and sounds

The office follows local time: late at night and in the evening the lawn and the street turn bluish and dark, the streetlights, the lamps, the machines and the monitors that are on cast halos of light, the rooms with people in them are lit (the lights that go out when a room empties still apply), the cars drive by with their headlights on and fireflies show up in the garden. At dawn (~5–7 a.m.) and at dusk (~5–7 p.m.) everything takes on a warm tone, and during the day the sun comes in through the windows and draws strips of light on the floor — short at noon, long and orange in the late afternoon. In Settings › Day/night cycle you can choose automatic, always day or always night; to test a time, use ?hora=21:30 in the URL.

Sounds come turned off. Turned on in Settings › Sounds, they're synthesized in the browser itself (no audio files) and quiet: the typing of whoever is working in the rooms in view, the elevator "ding" when someone arrives or leaves, the bell when someone needs you, the click of a finished task and the ping-pong and arcade sounds of the hangouts. There's a master volume and each category can be turned on and off separately. With the tab hidden, only the bell rings.

Terminal

Click an agent and use Open terminal to see the session's conversation the way Claude Code shows it: the prompts, the replies, each tool call (with the command or the diff) and the result, updated live. It works for main agents and subagents; in demo mode, the conversation is made up.

At the bottom, for main agents, you can type: Enter sends and Shift+Enter adds a new line. The text goes into the session as if you had typed it in its terminal (see Sending messages); if the agent is busy, it goes in when the agent finishes what it's doing. Subagents and sessions from the history stay read-only.

• Search: with the terminal focused, Ctrl+F (⌘F on Mac) or the magnifier in the header opens search in the conversation, ignoring case and accents. The counter shows the position ("3/17"); Enter and Shift+Enter go to the next and the previous one, expanding the collapsed blocks ("… +N lines") where the term is. Search keeps working while new messages arrive. Esc closes the search; the next one closes the terminal.
• Filter: All, Prompts only (your prompts and the agent's final replies, without the intermediate steps) or No tools.
• Copy: hover over (or focus) a prompt, a reply, a command or a result to copy it.
• Session history: the clock in the top bar lists the sessions from the last 7 days across all accounts (up to 150), grouped by day, with search by title, project or account. An ended session opens in the terminal with the project, title and date in the header and "Session ended at …" at the bottom; a session that's still open opens the agent's live terminal.

Since the terminal (and the history) shows the whole conversation, it only exists when Habblaud is reachable only from the computer it runs on (the default), and it only opens through http://localhost or http://127.0.0.1. With the port open to the network (HABBLAUD_BIND=0.0.0.0 or HABBLAUD_HOST=0.0.0.0), it stays off. Details in Privacy and security.

Interactive terminal (take over, new session and end)

With HABBLAUD_PTY=1 and running with npm start (outside Docker), each main agent's terminal window is interactive: the drawer's Open terminal button, the shortcut next to each agent in the side list or the T key open the same window, with tabs for the other agents in the same project.

• Session opened here: it's the real Claude Code (type, answer; Esc interrupts the agent, as in the terminal).
• Session opened in another terminal: Take over here (at the top of the window) ends the session there and continues the same conversation here (claude --resume <sessão>, in the same folder and account). The character keeps its name. On Windows, the terminal the session came from goes back to normal (Habblaud turns off the mouse/keyboard modes that Claude turned on).
• End closes the session, whether it's here or in another terminal (the conversation is saved).
• New session (in the room's drawer) opens a new Claude Code in the project folder.

Taking over and ending ask for a second click; closing the window doesn't end the session. The window (and the agent's drawer) can be dragged by its bar and resized from the corner; double-clicking the bar puts it back in place. It comes turned off: turn it on with HABBLAUD_PTY=1 (first read the warning in Privacy and security). It uses node-pty, @xterm/headless (optional dependencies) and the claude on your PATH (HABBLAUD_CLAUDE_BIN for another one); it only accepts requests from the computer it runs on. On startup, Habblaud opens a test terminal: if node-pty or claude don't work, the feature stays off and the log says why. Restarting Habblaud ends the sessions it opened. The live terminal is still available in the session history, for subagents and when the interactive one is off.

Day timelapse

The Timelapse button (a clock with a back arrow, or the L key) replays the day at high speed: rooms lighting up and going dark, agents arriving, working, waiting on you, heading to the hangouts, subagents coming and going. The playback bar has the day, play/pause, the speed (60×, 180× or 600×: a 10-hour day in 10, 3⅓ or 1 min), the draggable timeline with the chart of who was present and working, the peak markers (click to jump there) and Back to live. Meanwhile, the office takes on a light sepia tone, with the REPLAY 14:32 badge, and the feed keeps showing what's happening now.

The server records the timeline from the moment it's turned on (the past can't be rebuilt): summaries of what the office shows, without conversations, in ~/.habblaud/timeline/ (in Docker, in the data volume), with a size limit per day and the last 7 days kept. To not record: HABBLAUD_TIMELINE=0. Details in server/README.md.

The characters walk faster in the replay, but at high speeds anyone who stays only briefly in the office barely reaches their desk; jumping to another point shows everyone already in place.

Answering from the office

With the permissions plugin installed (npm run mod:install, which installs it along with the mod; on Claude Code older than 2.1.287, npm run hooks:install), when an agent asks for permission — to run a command, edit a file, open a page — the request shows up in the office: the character raises a hand, a notice with Reply pops up on screen and, in the agent's details, the Asks for permission card shows the command (or the edit's diff) with the buttons:

• Approve — and, when Claude Code suggests it, Approve and don't ask again (the same rule the terminal would offer, for example Bash(npm test:*) in this project);
• Deny — with an optional reason, which goes to the agent, and the option to interrupt it;
• Answer in terminal — Habblaud sets the request aside.

The dialog keeps showing in the terminal at the same time, and whichever answer comes first wins: if you answered in the terminal, the request disappears from the office on its own. The need you counter (and the P key) takes you to each request.

The agent's questions (AskUserQuestion) work too: the questions card shows each question with its options and descriptions, as the terminal would. Pick an option (or several, when the question allows it) or write your own in Other, and use Reply; Deny… tells the agent you don't want to answer, and Answer in terminal leaves the question there. The answer reaches the agent as if you had chosen it in the terminal.

• The request goes through a Claude Code PermissionRequest hook (the plugin's, or the old way's), which only diverts it when there's some Habblaud page open on this computer; with Habblaud stopped or no page open, it exits right away and the terminal carries on as usual.
• With no answer from the office within 5 minutes (up to 24 h with HABBLAUD_PERMISSION_TIMEOUT=86400 in the env of settings.json), the request goes back to being answerable only in the terminal (the plugin's options are in mod/README.md; in the old way, npm run hooks:install -- --timeout 120 changes the time and --port, the port). For background subagents, Claude Code only shows the dialog in the terminal after the hook finishes: answer from the office or use Answer in terminal. With a wait longer than 5 minutes, if no page that can answer stays open for 5 minutes (and no phone with login is set up to answer), the request goes back to the terminal.
• It works under the same safeguard as the terminal: only with Habblaud reachable only from the computer it runs on and opened through http://localhost. Check with npm run mod:status (or npm run hooks:status, in the old way). It also works from the phone with login (its default).
• In demo mode, the made-up agents also ask for permission and ask questions (pretend ones), so you can try it out.

Sending messages

With the messages plugin installed (npm run mod:install, which installs it along with the mod; it requires Claude Code 2.1.287 or newer), each main agent's details get the Send message section, and the bottom of the terminal becomes a text box. Write and press Enter (Shift+Enter adds a new line): the text goes into the session as if you had typed it in its terminal. You can reply when the agent finishes its turn and waits for the next request, give a new instruction or correct its course, without hunting for the right window.

• The box shows the status: queued (the session fetches messages every 2 seconds), delivered or not delivered, with the reason. With the agent busy, the message waits and goes in when it finishes its turn, like a prompt typed while the agent is working.
• The plugin runs inside each session and only talks to Habblaud at 127.0.0.1. Without it (or in a session opened before the install: use /reload-plugins), an install hint shows up in place of the box. Subagents don't receive messages.
• It works under the same safeguard as the terminal (local access only) and can be turned off with HABBLAUD_MENSAGENS=0. Read the warning in Privacy and security. From the phone with login, only with messages in HABBLAUD_REMOTE_ALLOW.
• In demo mode, the made-up agents also receive messages (pretend ones).

Editing the character

Each session arrives with a random name and look. To pin a project's character, open the main agent's details and click the pencil next to the name (Edit character). In the editor:

• Name: up to 24 characters. It can't repeat the name of someone who's in the office or the one chosen for another project.
• Randomize: brings up another look.
• Pieces: skin, hair, beard, eyes, top, bottom, shoes and accessory, with a preview next to them.
• Save: applies to the project (the room). The next session opened in it arrives as this character. With two sessions at the same time in the same room, the second one gets a random character.
• Back to random: undoes the choice.

The choice is stored in ~/.habblaud/names.json and goes away after 60 days without the project being used. The editor has the same safeguard as the terminal: the pencil only shows up with the terminal on (without HABBLAUD_TERMINAL=0) and local access, that is, with Habblaud reachable only from the computer it runs on and opened through http://localhost.

Codex

Codex sessions (CLI or app) show up like Claude Code's: character, project room, activity, tasks, subagents, terminal and history, 5-hour and weekly usage and My day (tokens only: Codex doesn't report cost). The Codex account chip is hollow and carries the CODEX tag.

• Without installing anything, Habblaud reads the files Codex itself writes (the conversations in ~/.codex/sessions and the locks of the open sessions in ~/.codex/thread-writer-locks). The catch is that Codex writes each step after it finishes and never writes "waiting on you": without the hooks, the office shows Codex with a delay and doesn't tell you when it needs you.
• With the hooks (npm run codex:install and approving them in /hooks), Codex notifies Habblaud at every step: the session shows up right away, the activity is the current one and an approval request makes the character raise a hand.
• Approving from the office: with some Habblaud page open, Codex's approval request waits for your answer in the Asks for permission card for up to 25 seconds (npm run codex:install -- --espera <s>, from 5 to 120): Approve or Deny (with the reason). Unlike Claude Code, Codex only shows the approval in the terminal after the office answers or the time runs out — in the meantime the terminal shows "Waiting for an answer in Habblaud…". There's no "always allow" or "interrupt", and Codex doesn't let you answer its questions from outside.
• Sending messages: the Send message box works for Codex too, through codex queue: the message joins the session's queue and Codex picks it up when it finishes what it's doing (up to about 10 seconds later). In Node mode Habblaud runs the command itself; in Docker, keep npm run codex:bridge running on the Mac (the container can't see your Codex).
• 5-hour and weekly usage: comes from Codex's own files and only refreshes while some session is running; the card shows how old the numbers are and "no quota" when the workspace has run out of credits.
• When the session comes in: with the first message (that's when Codex records the project folder); a CLI that's open and still has no conversation doesn't show up.
• When the session leaves the office: the Codex CLI runs sessions on a background server, which keeps them loaded for up to a minute after they go idle with no one watching; that's why the character may take a little while to leave after you close the terminal.
• ChatGPT cloud conversations (with no file on the computer) don't show up. HABBLAUD_CODEX=0 turns Codex off.

GitHub in the office

What the agents do on GitHub livens up the project's room, with no token and no internet access: Habblaud reads the calls in the transcripts (gh pr create, gh pr merge, git push, gh run watch, gh pr checks, gh run view, gh release create and the GitHub MCP tools) and their results.

Event	What happens
🎉 PR opened or merged, release published	Confetti falls in the room, everyone celebrates by hopping and a banner says why ("PR #12 merged!") for ~12 s.
🚨 Red CI	A rotating beacon flashing in the corners of the room, a reddish floor and a "!" bubble over whoever saw the failure, until a green CI in the room (which turns into a party) or for 10 min.
🚀 Push	Just the notice and the feed.

Each event also creates a notice ("🎉 Danilo opened PR #12 in habblaud", "🚨 CI failed in habblaud (feat/x)") and goes into the feed. Only what happens live livens up the room: whatever was already in the transcripts when Habblaud opened only goes to the history. With "reduce motion" turned on in the system, nothing flashes or spins. In demo mode, made-up PRs, merges and CI runs show up from time to time.

My day

The My day button (or the M key) shows where the day went: how much time the agents spent working and, highlighted, how much time they spent waiting on you (permission, question or choice) — with the longest wait, who waited, where and when. Next to it, sessions, subagents, requests, tokens and cost (when Claude Code records the cost in the transcript), and charts by hour, by project (from the one that waited on you the most to the one that waited the least) and by account, plus the ranking of the longest waits. Pick the day in the selector (Habblaud keeps the last 30); when it's showing today, the panel refreshes every 30 s. Each chart has a table version, and the legend turns each status on and off.

The time is agent time: two agents working for an hour count as two hours. Habblaud only counts what happens while it's running, and the day follows your browser's time zone. With demo mode on, the panel shows made-up numbers (with the "demo" badge) and lets you switch to the real data; the demo agents never go into the real stats.

Demo mode

Want to see the office full without having any sessions open?

• Add ?mock=1 to the URL (e.g. http://localhost:4747/?mock=1) to simulate everything in the browser, without your real sessions. Parameters: &speed=3 speeds up time and &sessions=6 changes the number of simulated sessions.
• Or turn it on from the interface — View demo (when the office is empty) or Settings › Demo mode — which puts made-up agents alongside the real ones. With npm run demo (or HABBLAUD_DEMO=1 in the Docker .env), this mode starts already on.

Accounts and usage (5-hour and weekly)

Codex accounts come in on their own (each ~/.codex* folder or CODEX_HOME is an account; the letter comes from an alias x='CODEX_HOME=… codex', if there is one) and their usage comes from Codex's files; see Codex. The rest of this section is about Claude Code.

If you use more than one Claude Code account (for example, a c shortcut with ~/.claude and a d with CLAUDE_CONFIG_DIR=~/.claude-conta2), Habblaud shows them all together: each agent carries its account's chip and the top bar shows, for each account, the percentage used of the 5-hour session and of the week and when each limit resets.

How accounts are recognized. Each Claude Code config folder is an account: ~/.claude and any ~/.claude* with projects/ or sessions/ (plus CLAUDE_CONFIG_DIR). To give each one its letter, Habblaud reads, in your shell files (~/.zshrc, ~/.bashrc…), only the alias x='... claude ...' lines:

alias c='claude'
alias d='CLAUDE_CONFIG_DIR=~/.claude-conta2 claude'

becomes Account C and Account D. Only shortcuts of up to 3 letters count; accounts without a shortcut get A, B…

Where the numbers come from. The recommended way is the Habblaud mod (npm run mod:install, Claude Code 2.1.287+): inside each session, it receives the plan limits from Claude Code itself and stores only the 5-hour and weekly percentages (and when they reset) in ~/.habblaud/usage/<conta>.json. The numbers arrive after each account's next reply. Check with npm run mod:status. On earlier versions of Claude Code, the statusline tap (npm run usage:install) does the same through the statusline command, in the same file.

• usage:install (the old way) changes, in <conta>/settings.json, only the statusLine.command field: the original command (e.g. npx -y ccstatusline) now runs through node "<pasta do Habblaud>/scripts/statusline-tap.mjs" -- <comando original> (on Windows, the original command runs in Git Bash, as Claude Code would run it). A copy of the file first goes to settings.json.habblaud-backup-<data>. If the account had no statusline, one is created that only captures usage. npm run usage:install -- --dry-run shows what would change without writing anything.
• The tap passes the same JSON on to your statusline (output and exit code are still its own) and writes only {accountId, configDir, fetchedAt, five_hour, seven_day} — no prompts, costs or project paths. Any failure while capturing is ignored: the statusline never breaks because of Habblaud.

Source	How it works
Habblaud mod (recommended)	Live, as described above. Replaces the tap: mod:install removes the tap from the account.
Statusline tap (Claude Code older than 2.1.287)	Live, through the statusline command, in the same file.
/usage cache (always on)	Claude Code saves the last result of the /usage command. In Node mode, Habblaud rereads it every 60 s; in Docker, the value read at the last npm run docker:up applies. It only changes when someone runs /usage.

The sources are local files: none of them reads passwords or tokens, or makes network calls. The source with the most recent numbers always wins. Numbers older than 30 minutes show up as outdated; a window that has already reset since the numbers were collected shows up as — until new numbers arrive.

Configuration

Everything works without configuring anything. If you need to adjust something, use environment variables:

Variable	Default	What for
HABBLAUD_PORT	4747	HTTP port (in Docker, the port published on the host).
HABBLAUD_HOST	127.0.0.1	Server interface in Node mode. Anything other than 127.0.0.1/localhost turns off the terminal (and answering and sending messages from the office).
HABBLAUD_BIND	127.0.0.1	Docker only (in .env): where the port is published. 0.0.0.0 opens it to the local network (and turns off the terminal and anything that acts on the sessions).
HABBLAUD_CLAUDE_DIRS	automatic detection	Account folders, comma-separated (e.g. /caminho/conta1,/caminho/conta2).
HABBLAUD_DATA_DIR	~/.habblaud	Where Habblaud keeps its own data (character and room names, the timelapse timeline and the My day stats).
HABBLAUD_TIMELINE	on	0 turns off timeline recording (the days already recorded stay in the timelapse). In Docker it's always on.
HABBLAUD_USAGE_DIR	~/.habblaud/usage	Where the mod (or the statusline tap) writes usage.
HABBLAUD_DEMO	off	1 turns on demo mode at startup.
HABBLAUD_ALLOWED_HOSTS	—	Extra names accepted in the address (e.g. meu-mac.local), besides localhost and IPs.
HABBLAUD_TERMINAL	on (local access only)	0 turns off the terminal, answering and sending messages from the office, and the character editor. With the port exposed they're already off, with no way to turn them on.
HABBLAUD_PTY	off	1 turns on the interactive terminal (open, take over and end Claude Code sessions from the office; local access only and outside Docker) and New agent (also in Docker, through npm run codex:bridge). Read the warning in Privacy and security.

In Docker, HABBLAUD_PORT, HABBLAUD_BIND, HABBLAUD_ALLOWED_HOSTS, HABBLAUD_DEMO, HABBLAUD_TERMINAL, HABBLAUD_MENSAGENS, HABBLAUD_CODEX, HABBLAUD_UPDATE_CHECK, HABBLAUD_LANG, HABBLAUD_PTY, the HABBLAUD_WORLD* and the HABBLAUD_REMOTE_* apply (in .env or in the environment), as do HABBLAUD_CLAUDE_DIRS, HABBLAUD_CODEX_DIRS, HABBLAUD_USAGE_DIR and HABBLAUD_ACCOUNTS (read by docker:up on the host); the rest are fixed inside the container. Options: npm run docker:up -- --no-build (starts without rebuilding), npm run docker:down (stops) and npm run docker:logs (follows the logs).

Habblaud World

Habblaud World is an online pixel-art city where every Habblaud becomes a building, lit up live while the agents work, and everyone walks around with a character. It's on by default, pointing at server.habblaud.com; nothing leaves your computer until you create an account and sign in, and HABBLAUD_WORLD=0 turns everything off. Access is granted through a waiting list.

How to configure it. The variables below (in .env or in the environment) are optional: without them, Habblaud points at the official server. The section shows up in the Habblaud World window (the city button in the top bar) and only works from this computer (never through the phone port); the account stays only here, in ~/.habblaud/world.json (permission 600; in Docker, in the /data volume).

Variable	Default	What for
HABBLAUD_WORLD_URL	https://server.habblaud.com	World server address, https:// or wss:// only (no path).
HABBLAUD_WORLD	on	0 turns off everything about the World (routes and interface) and nothing leaves your computer.
HABBLAUD_WORLD_DEV	off	1 accepts a local world server without TLS: http://127.0.0.1:<port> (npm run dev already does) or, from inside Docker, http://host.docker.internal:<port>.

Privacy. Nothing about your work leaves your computer. From the building, only this goes out: the theme, one floor per open room (with a random code that changes every day) and, for each agent, main or subagent, the tool, the status, the appearance, a drawn name and the band of tasks completed that day. Paths, project names, titles, commands, messages, code, cost, model, branch and your Claude account never leave. The full list, field by field (in Portuguese): docs/world/privacidade.md. See what was sent, on the Your building card, shows the exact JSON of the last upload, with its size in bytes and the world's reply.

Account and waitlist. Sign up with e-mail, player name (3 to 20 characters, unique in the world: it stays reserved for you while you wait on the list and becomes your character's name) and password (12 characters or more; leaked passwords are refused), a 6-digit code by e-mail and accepting the Terms of Use and the Privacy Policy. Every account joins a waitlist, and access is granted in batches: an open Habblaud checks on its own every 30 minutes. Up to 3 computers per account (Devices, with revoke).

Character. The player name chosen at sign-up (it can be changed later, once every 30 days), the same appearance editor as the agents, 2 or 3 personality traits and an intro line from the list (there's no free text in the world besides the name). You can create it while still on the waitlist.

Building. With the account approved and the character created, the building goes up on its own (over a WebSocket) and follows the office live. The Your building card shows the connection, floors, agents, the last upload and, when the world refuses an upload or stops the building (old version, another computer, suspended or banned account), the reason; a suspended or banned account gets Appeal by e-mail. A building marked as "unverified" (implausible numbers, under review) stays out of leaderboards and highlights and shows Dispute by e-mail. When the world is under maintenance, the card says so, and Habblaud tries again on its own.

Visibility. Live (floors, counters and each character's status) or discreet mode (only lit or dark and the number of floors); closed to visits; hide the country; the building's tokens only in bands (up to 100 thousand, 1 million…), never the number or the cost, and they can be turned off; leaderboard and compactions only if you turn them on; knocks on the door (Let me know when someone knocks) only if you turn it on.

City. The World button in the top bar swaps the office for the city: walk with the arrows/WASD or by clicking, each building's card shows floors, agents and status, and the minimap sits in the corner; paused or out of the world, your building stays dark, but you can still walk around the city. In demo mode (?mock=1&world=approved&scene=city, only fictional buildings and people): the city and a building's card.

Visits. Near a building's door, Enter (E) takes you inside: the owner's live office, with rooms under made-up names and the agents in their current status, view only (no terminal, messages or permissions, and nothing about the work); a building closed to visits can only be seen from outside. On yours, you see "N visitors now" (only the count, never who). In demo mode: the visit and the visit panel.

Social. In the city, with no free text anywhere: L likes the building by the door, G opens the greeting wheel (ready-made phrases, translated for whoever reads them) and K knocks on the door of owners who accept it; each person's card has Greet, Block and Report, and the minimap takes you to a building by its code ("Go to HB-XXXX") or to any lit building. Inside a building, Like and Leave a stamp at the front desk (one per building per day). In the square, the thermometer of the city's weekly goal and, when it's reached, fireworks. In the office, the "❤️ N today" chip and the "While you were away" card with the visits, likes and stamps left at the front desk (never who); people you blocked are listed in Blocked, with Unblock. In demo mode: the greeting wheel and the square with the fireworks.

Pause, leave, delete. Pause turns the building off right away and keeps the lot (after 30 days paused, it disappears from the map; 30 days later, it's removed). Leave the world removes the building and frees the lot; the account and the character stay, and coming back creates a new building. Delete account (in Account and data, with the code that arrives by e-mail) takes the building and the profile out of the world right away, deletes the account, likes, blocks, reports you made and devices from the database within 7 days (and from backups within 30) and deletes world.json from this computer. With no word from Habblaud for 7 days, the building disappears from the map and, 30 days later, is removed; an e-mail warns 3 days before (can be turned off in Account and data).

Export. Download my data (in Account and data) saves habblaud-world-dados.json with what the world keeps about your account: e-mail, dates, character, visibility, devices, building, likes given, blocks and reports you made. Once a day.

Truthful data. The World only works if the buildings are real: changing Habblaud to send invented or inflated floors, agents, statuses, tokens or characters breaks the Terms of Use and gets the account banned (the server checks the plausibility of every packet). Changing Habblaud for anything else stays free (MIT). How Habblaud talks to the World (in Portuguese): shared/world/README.md.

Testing with a local world server. For World development, with the world server running on the same computer (for example, in Docker, published on port 4848):

• Outside Docker: HABBLAUD_WORLD_URL=http://127.0.0.1:4848 npm run dev.
• With Habblaud in Docker: in this folder's .env, HABBLAUD_WORLD_URL=http://host.docker.internal:4848 and HABBLAUD_WORLD_DEV=1; then npm run docker:up, which ends with the World line (Habblaud World: on: http://host.docker.internal:4848, or why it stayed off). docker-compose.yml already maps host.docker.internal to the host (extra_hosts with host-gateway; Docker Desktop already has the name). On Linux, the world server must listen on an interface the container can reach (0.0.0.0 or the Docker bridge IP): a port published only on the host's 127.0.0.1 doesn't answer through host-gateway.
• Without TLS, only with HABBLAUD_WORLD_DEV=1 (or --dev) and only for 127.0.0.1 and host.docker.internal; otherwise, only https:// or wss://. The world server must accept the name used in the URL as its game host (otherwise it answers 421).

How the section looks, in demo mode (?mock=1&world=logged_out, waiting or approved; in Portuguese): no account, waitlist, character, Your building and what was sent.

How it works

 ~/.claude*/sessions/<pid>.json ─────────────────┐
 ~/.claude*/projects/<projeto>/<sessão>.jsonl ───┼──▶ servidor Node ──▶ modelo do escritório ──▶ SSE ──▶ navegador
 .../<sessão>/subagents/**/agent-*.jsonl ────────┘    (polling de 1 s)                                 (canvas + interface)

• Who's in the office: while a session is open, Claude Code keeps sessions/<pid>.json with the project and the status (busy, idle, waiting on you or waiting on a background shell). That's what decides who shows up and what each one does.
• Running shells: in the transcripts, each background Bash (or long foreground command) becomes a "shell" with a label and a timer, until the notification arrives that it finished, failed or was interrupted.
• Social life: the hangouts, the personalities and the coins are a simulation in the browser (client/src/world/social/), with no effect at all on the real agents. The personality comes from each agent's seed (it's the same in any browser); the wallets are kept in the browser's localStorage.
• What each one is doing: the server follows the end of the transcripts (.jsonl) of the open sessions and turns each tool call into an activity, with an icon, in the server's language (see Languages). Tasks, title, model and stats come from there too.
• Subagents: the transcripts in <sessão>/subagents/ (including those from workflows) become characters linked to the agent that called them.
• GitHub: PRs, pushes, CI and releases come from the same transcript lines (the command and its output, or the gitOperation that Claude Code itself records); nothing is queried on GitHub.
• The drawing: the browser receives the state over SSE and draws everything on a canvas — characters, furniture, floors and walls are pixel art generated by code; the logo, the illustrations and the pictures on the walls, on the other hand, were generated with AI.

shared/   protocolo (types.ts), atividades em PT-BR, nomes, simulador de demonstração
server/   servidor HTTP + SSE: contas e uso, leitura das sessões e transcripts, modelo do escritório
client/   Vite: src/art (pixel art procedural), src/world (o escritório no canvas), src/ui (interface)
scripts/  build do servidor, docker-up, instaladores (mod, tap de statusline e hook de permissão) e screenshots
mod/      o mod do Habblaud e os plugins de permissões e de mensagens (plugins do Claude Code; marketplace em .claude-plugin/)

Route	Description
GET /api/stream	SSE with the snapshot, feed and notice events (format in shared/types.ts).
GET /api/snapshot	Current state of the office.
GET /api/agents/:id	An agent's details, with up to 200 activities.
GET /api/agents/:id/terminal	Terminal SSE (init and append events); local access only.
PUT /api/agents/:id/character	Edit the character: {name, seed, parts} saves the main agent's name and look as the project's (the room's) character; local access only.
DELETE /api/agents/:id/character	"Back to random": deletes the project's character and the agent goes back to its random name; local access only.
GET /api/sessions/recent	Session history: sessions from the last 7 days across all accounts (up to 150); local access only.
GET /api/sessions/:conta/:sessionId/terminal	SSE of the conversation of a session from the history (same protocol as the terminal); local access only.
GET /api/stats?day=AAAA-MM-DD	My day stats (time per status, projects, accounts, hours, waits, tokens and cost).
GET /api/stats/days	Days with stats (the last 30).
GET /api/timeline/days	Days recorded for the timelapse, with size and the time of the first and last entries.
GET /api/timeline/:dia	A day's timeline (AAAA-MM-DD), in JSONL (gzipped).
GET /api/health	Health: version, demo, Docker, terminal, answering and sending messages from the office, sources and usage status of each account.
POST /api/demo	{"enabled": true | false} turns the simulated agents on or off.
POST /api/rooms/rename	{"id": sala, "name": nome} renames the room (empty goes back to the folder name); the name is stored in rooms.json, in the data folder. Local access only.

More server details in server/README.md.

Host	Container	What for
<conta>/sessions/	/claude/<conta>/sessions (read-only)	Open sessions and their status.
<conta>/projects/	/claude/<conta>/projects (read-only)	Transcripts of the sessions and subagents.
~/.habblaud/usage/	/usage (read-only)	Usage captured by the mod (or by the statusline tap).
habblaud-data volume	/data	Habblaud's own data (character names, the timelapse timeline and the My day stats).
~/.habblaud/remote-tls/server/	/remote-tls (read-only)	Only after npm run remote:setup: the server certificate and key of the phone with login (never the CA key).

The account folder is never mounted whole (credentials and settings live there). The container runs as an unprivileged user, with a read-only file system, no extra capabilities and with no-new-privileges. The accounts' metadata (letter, email, organization) is read on the host by docker:up and passed to the container.

Privacy and security

• Read-only: Habblaud never writes to the Claude Code folders. The exceptions are the installers, which you run: npm run mod:install / mod:uninstall uses claude plugin itself (which registers the marketplace and the plugins in settings.json and in <conta>/plugins/) and, on install, removes the old tap and hook with a backup; in the old way, npm run usage:install / usage:uninstall changes only the statusLine.command in settings.json and npm run hooks:install / hooks:uninstall, only the hooks.PermissionRequest list — always with a backup first.
• The mod: runs inside each Claude Code session, with your permissions (like every mod), and only does what is listed in mod/README.md — claude plugin validate mod/habblaud shows the events and the calls, without running anything. It only writes the 5-hour and weekly usage percentages (and when they reset) to ~/.habblaud/usage/<conta>.json; for the "needs you" line and /habblaud, it only talks to Habblaud at 127.0.0.1. It doesn't read the conversation, doesn't call the model and doesn't send anything off the computer. It's loaded from this folder: whatever is in it (including after a git pull) is what runs. The messages plugin (habblaud-mensagens) is separate precisely because it acts: it types into the session what you sent from the office.
• Local only, by default: the server only accepts connections from the computer it runs on; opening it to the local network is optional. There's no telemetry. The only external call is the new-version check: every 6 hours, an anonymous request, with no token, to the public GitHub API (api.github.com/repos/marmottajr/habblaud/releases/latest). It doesn't send anything about your sessions; GitHub only sees your IP and the version in use, which goes in the User-Agent. HABBLAUD_UPDATE_CHECK=0 turns off the request. Habblaud World (on by default) only sends something to the world server after you create an account and sign in, and only what is listed in what leaves your computer (in Portuguese); turns it all off. So is the , which only exists after .

Development

Command	What it does
npm run dev	Server + Vite at http://localhost:4747, with hot reload.
npm run dev:client	Vite only, at http://localhost:5173 (open it with ?mock=1; /api is proxied to port 4747).
npm run demo	npm run dev with demo mode on.
npm run build / npm start	Builds and runs the production version.
npm test	Tests (Vitest).
npm run typecheck	Type checking for the client, the server and the scripts.
claude plugin test mod/habblaud	Mod tests (no session, account or network); same for mod/habblaud-mensagens.
claude plugin validate .	Validates the marketplace (.claude-plugin/marketplace.json); with mod/habblaud, mod/habblaud-permissoes or mod/habblaud-mensagens, validates the plugin and lists the mod's events and calls.
claude --plugin-dir mod/habblaud	Opens a session with the mod from this folder, without installing it (edit and run /reload-plugins).
npm run mod:status	Shows, per account, the marketplace, the plugins and the installed versions.
npm run demo:timeline	Generates a made-up timeline (demo mode simulator) for the timelapse.
npm run release	Publishes the package.json version on GitHub (tag and release with the notes from CHANGELOG.md).

To take screenshots without opening your browser, scripts/shot.mjs uses an isolated headless Chromium (Playwright). The first time, download the browser (just once): npx playwright-core install chromium-headless-shell.

node scripts/shot.mjs 'http://localhost:4747/?mock=1&speed=3' /tmp/habblaud.png --size 1400x900 --wait 4000

In the browser console, habblaud.world.debug has tools for testing scenes (e.g. habblaud.world.debug.setHour(21) to see the night). To set the time right when it opens, use ?hora=21:30 in the URL.

Timelapse with made-up data only (e.g. to record a GIF): npm run demo:timeline runs the demo mode simulator offline and records a whole day (by default, yesterday from 9 a.m. to 7 p.m., with a busy morning, a quieter lunch and a peak in the afternoon) in <tmp>/habblaud-demo/timeline/. Then start a Habblaud that only reads that folder and doesn't show your sessions, and open the Timelapse:

npm run demo:timeline -- --data-dir /tmp/habblaud-demo        # opções: --date, --start 8:30, --hours, --sessions, --seed
mkdir -p /tmp/habblaud-demo/vazio
HABBLAUD_DATA_DIR=/tmp/habblaud-demo HABBLAUD_TIMELINE=0 HABBLAUD_CLAUDE_DIRS=/tmp/habblaud-demo/vazio \
  HABBLAUD_PORT=4848 npm start                                 # depois de npm run build; abra http://localhost:4848

HABBLAUD_TIMELINE=0 avoids writing to that folder; with recording on, days older than 7 days are deleted.

Publishing a version: every version has its own section in CHANGELOG.md, with what went in. While working, note the changes under "Não lançado" (Unreleased). To release:

1. Bump the number with npm version minor --no-git-tag-version (or patch, for a bug-fix-only version).
2. Replace "Não lançado" with the version and the date (## [0.3.0] - AAAA-MM-DD), leave an empty "Não lançado" above it and update the links at the end. npm test fails if the package.json version doesn't have a section.
3. Merge into main and run npm run release (it needs an authenticated gh; -- --dry-run only shows the notes). The script creates the v0.3.0 tag and the release with the section's text. From then on, Habblaud users see the New version badge.

Troubleshooting

Only open Claude Code sessions show up. Check whether there's one running and whether the account was detected at http://localhost:4747/api/health. If the account folders are somewhere else, use HABBLAUD_CLAUDE_DIRS=/caminho/conta1,/caminho/conta2.

Run npm run mod:install (or, on Claude Code older than 2.1.287, npm run usage:install) and check with npm run mod:status (or npm run usage:status). The numbers arrive after the next reply in an open session of that account; in a session opened before the install, run /reload-plugins or reopen the session. A — in place of the percentage means the window has reset since the last collection.

Run npm run mod:status: it shows the Claude Code version (the mod needs 2.1.287 or newer) and, per account, whether the marketplace points to this folder and whether the plugins are installed and enabled. In a session opened before the install, run /reload-plugins; /plugin shows, below the tabs, the loaded mods (e.g. 1 mod active · habblaud). Mods don't run with "disableAllHooks": true in the account's settings.json or with claude --safe-mode.

Another Habblaud (Node or Docker) is already running. Stop it (npm run docker:down, or Ctrl+C in the npm start terminal) or use another port: HABBLAUD_PORT=4848 npm run docker:up (Docker) or HABBLAUD_PORT=4848 npm start (Node).

Check that you created .env with HABBLAUD_BIND=0.0.0.0 and recreated the container (npm run docker:up -- --no-build), that the phone is on the same Wi-Fi and that the computer's firewall allows connections on port 4747. By IP it works right away; to open it by a name (e.g. meu-mac.local), add HABBLAUD_ALLOWED_HOSTS=meu-mac.local (in Docker, in the same .env).

The terminal only exists with Habblaud reachable only from the computer it runs on. Check that .env doesn't have HABBLAUD_BIND=0.0.0.0 (or, in Node mode, that you didn't use HABBLAUD_HOST=0.0.0.0) or HABBLAUD_TERMINAL=0, and open it through http://localhost:4747 (through the network IP or a name from HABBLAUD_ALLOWED_HOSTS it's refused). In Docker, a container created before this feature needs to be recreated: npm run docker:up. The startup log (npm run docker:logs) says whether the terminal is on and, if it isn't, why.

Run npm run mod:status: it says, per account, whether the habblaud-permissoes plugin is installed and enabled (and whether the old hook was left alongside it, which would make two of them answer) and whether Habblaud is handling requests. In the old way, npm run hooks:status says whether the hook is installed and pointing to this folder. The request is only diverted with some Habblaud page opened through http://localhost (or 127.0.0.1) and with the terminal on (same safeguard). Sessions opened before the install load the plugin with /reload-plugins (or when reopened). In the old way, if Habblaud uses another port, reinstall with npm run hooks:install -- --port <porta>. For the agent's questions, the plugin must be from 0.6.0 or newer: run npm run mod:install after updating.

The startup log (npm run docker:logs) says how many Codex accounts Habblaud found; if it's none, point to the folder with HABBLAUD_CODEX_DIRS (and run npm run docker:up again: in Docker, only the folders that existed at docker:up are mounted). Without the hooks, Codex shows up through its files, which are only written after each step: run npm run codex:install, approve the hooks in /hooks in Codex and check with npm run codex:status. Codex runs the hooks through the login shell, which may have an old Node; codex:install picks a Node 22+ and tells you which one (or use --node <caminho>). A freshly opened CLI session only shows up with the first message: before that, Codex doesn't say which folder it's in (it only creates the session lock), and without the folder there's no room.

The box only shows up for main agents whose session has the habblaud-mensagens plugin connected: run npm run mod:status (it says, per account, whether the plugin is installed and whether Habblaud has messages turned on) and, in sessions opened before the install, /reload-plugins. Without the plugin, the agent's details show the install hint. Messages follow the terminal's safeguard and HABBLAUD_MENSAGENS=0 turns them off; the startup log (npm run docker:logs) says whether they're on and, if they aren't, why. A message that stays queued and fails with "the session didn't fetch the message" means the session stopped asking Habblaud (it closed, or the plugin was turned off).

This happens if Claude Code was force-quit without deleting its own entry in sessions/ (in Docker there's no way to check whether the process still exists). Running with Node (npm start), Habblaud checks the processes and this doesn't happen.

Docker Desktop only shares some Mac folders with containers (by default /Users, /Volumes, /private and /tmp). Add the folder in Settings › Resources › File sharing.

npm run docker:down, then docker volume rm habblaud_habblaud-data (deletes the saved names) and npm run docker:up.

Disclaimer

Habblaud is an independent project, not affiliated with, sponsored or endorsed by Anthropic. "Claude" and "Claude Code" are trademarks of Anthropic, used here only to describe what the project works with.

To show the agents, Habblaud only reads local files that Claude Code itself writes on your machine (records of the open sessions, transcripts and account settings). This format isn't documented and may change between Claude Code versions: after an update, Habblaud may stop working, fully or partially, until it's adjusted.

License

MIT. The images in client/public/assets (logo, illustrations, pictures and posters) were generated with AI for this project and converted into pixel art by the scripts in scripts/assets/.

Social life

When two or more agents have nothing to do — idle, or waiting on a shell for more than 40 seconds — they stop wandering around alone and get together. Nobody plans anything: each character has a personality, friends and rivals, and that decides what it prefers to do, with whom and what it says. Someone waiting on a shell only leaves the desk for a hangout (alone, it sticks with the popcorn) and takes the hourglass along; when the command finishes — or when you send a new request to someone — it lets the group know ("Oops, they're calling me! 🏃") and runs back to its desk.

The hangouts

	Hangout	What happens
📺	TV in the lounge	Soccer, a soap opera or cartoons, with popcorn on the couch. During soccer the fans cheer along with the "GOL" flashing on the screen (or groan at the other team's goal); during the soap opera, little hearts and "No way! 😱". Nap lovers end up falling asleep on the couch.
🎮	Video games	Two on the couch with controllers, the TV in split screen (racing or fighting), best of 3, trash talk ("What's with this lag?!") and fans.
👾	Arcade	A duel at the lounge's two machines, a single match.
🏓	Ping-pong	Up to 5 points, with the ball going back and forth, the score in the bubbles ("3 × 2 🏓") and fans behind the table.
☕	Kitchen chat	Each one grabs a coffee (or water) and sits at the table: gossip with coworkers' names ("Did you see Rafaela's commit? 👀"), jokes (the group cracks up), work talk ("Deploy on a Friday? 😈") and topics about the day and the time ("TGIF! 🎉").
💬	Chat	Two coworkers standing in a corner of the lobby, the kitchen or the lounge.
✊	Rock-paper-scissors for coins	"Jo… Ken… Po!", the hand signs appear over their heads, the winner celebrates and the coins fly from one to the other; the loser gets upset and, if competitive, asks for a rematch. A tie goes again. Anyone walking by may stop to watch.
💄	Bathroom mirror	Lipstick or a comb, with the face reflected in the mirror and a sparkle at the end. Works alone or in pairs ("Can I borrow your comb?").

Personalities

Each agent has 2 or 3 traits, drawn from its seed (it's always the same person, in any browser):

🏆 Competition · 🎲 Betting · 🎮 Games · 📺 TV & shows · 🗣️ Gossip · 💄 Vanity · ☕ Caffeine · 😂 Jokes · 🏓 Sports · 🙈 Shyness · 💰 Thrift · 😴 Naps · 🧘 Calm · 📚 Reading

The traits change what each one chooses (TV lovers turn on the TV, sports lovers call people over for ping-pong), whom they call (friends look for each other; rivals challenge each other in games), what they say and even who plays better. Each one also has a catchphrase ("Done and done!", "Let's code!"). Hover over a trait in the agent's details to see what it does.

Coins

Made-up coins (🪙), just to make the bets more fun:

Arrived at the office	🪙100 (subagents, 🪙30)
Task done (item on the task list)	+🪙10
Request handled (end of turn)	+🪙5
Subagent delivered the result	+🪙15
Bets (rock-paper-scissors; matches between rivals or with betting lovers)	whatever the loser loses goes to the winner

Each gain shows up as "+🪙10" over the head; matches and bets go into the activity feed. The balance, the score and the statement are saved in the browser (each browser has its own economy: the bets are drawn there) and are forgotten three days after the agent leaves.

All of this is just Habblaud putting on a show: nothing changes in the Claude Code agents, which keep working normally.

---