DeckHQ GitHub

ComingNew

The same office, in 3D.

See the 3D office Tell me when it is ready

The same office twice: the DeckHQ floor seen from above, labelled Today, and beside it the same rooms in 3D, labelled Coming.
ComingThe same floor
A project room in 3D: two robots typing at a shared desk with code on their screens, and two juniors on the floor with laptops. A label reads Coming.
ComingA desk, close
One robot in each column of the picture, a state each: working at a desk in green, waiting for review in red, a hand up in amber, and powered down. A label reads Coming.
ComingEvery state
The office in 3D by night: one robot at a lit desk and a row of juniors on the floor behind it, their visors glowing green. A label reads Coming.
ComingThe night shift

An office for your AI coding agents

The agent that finished an hour ago is still waiting for you.

DeckHQ puts every Claude Code and Codex session on one office floor and keeps each one in your queue until you clear it.

npx deckhq app

Or count yours first

Node 18 or newer. Free, MIT, zero dependencies, local. Claude Code is verified; Codex is verified for reading and replying.

The DeckHQ floor, moving. A robot named Elif finishes its turn at a desk in the orbital-api room and a red badge appears over it. It stands, the room's lights go down behind it, and it walks the corridor into Your Office and sits on the sofa beside the two already waiting. The count in the header of Your Office goes from 2 waiting to 3.
One real walk on the demo floor, recorded at 60 frames a second. Elif finishes in orbital-api, crosses the corridor and sits in Your Office; its header goes from 2 waiting to 3.

Everyone waiting on you is in your office.

A session that finishes its turn stands up, walks the corridor into Your Office and sits down. How long it has waited is over its head, and the longest wait sits nearest your desk.

Two signals, because they need different answers.

  • A raised hand means I am mid-task and blocked. It needs a decision.
  • A red badge with a time means I finished; review this. It needs a read.
  • They are counted separately. The header has a number for each: 2 hands up, 4 for review.
Your Office with six robots waiting. Four sit under red badges reading 4h, 40m, 7m and 1d 2h, each with a tick. Two more, Zola and Kobe, hold a hand up. The room's header counts 6 waiting, 2 with a hand up and 4 for review.

Reading it does not clear it. Only you do.

What a session is doing changes on its own. What you owe it changes when you press a button. Opening a conversation does not clear it; scrolling past it does not clear it; reading it does not clear it.

Click anyone and the panel has how long they have waited, what they said as the markdown they wrote, and what changed in that project's working tree. Then 1 reply, 2 approve, 3 bench.

Go to session brings the terminal, the editor or the Claude desktop app that session is running in to the front.

The panel, in full

The session panel for a robot named Emeka, for review in orbital-api, waiting 1d 2h. It shows what the session said, a code block reading npm test, 214 passing, what changed in orbital-api at plus 142 and minus 18 across 3 files, and three buttons: 1 Reply, 2 Approve, 3 Bench.

A session that finished is still on the floor.

DeckHQ reads every transcript on disk. The session that finished an hour ago is still here, with what it last said, and it did not have to be started from DeckHQ.

The demo floor, moving: five project rooms with robots typing at desks, six robots waiting in Your Office with their waiting times over their heads, and fifteen resting in the lounge.
The demo floor: five repositories, twenty-one sessions. Every picture on this site is the running product, on invented projects.

One room per repository. One robot per session.

Start your agents anywhere: a terminal, your editor, the desktop app. Each one turns up in its project's room with a first name it keeps for good.

Every room has a header: its name, and four numbers in the same four places in every room. Here, 2 need you, 1 is working, none stalled, and 5.8M tokens today. Point at it and it says the same in a sentence.

One project room, orbital-api. Its header has four counts, each beside a small symbol: 2 need you, 1 working, 0 stalled, 5.8M tokens today. A green robot named Elif sits at the desk with a terminal bubble over its head.

Why not the agent view you already have?

Claude Code has one built in: claude agents. It starts sessions, keeps them running in the background and lets you answer one from a list. Use both. They do different jobs.

Agent view, in Claude Code DeckHQ
Which sessions Every background session you have started, across your projects. A session open in another terminal is listed once you send it to the background. Every transcript on your disk, wherever the session was started: a terminal, your editor, the desktop app.
For review A session with an open pull request that needs review, or with failing checks, moves to Ready for review. Every session that finishes its turn waits in Your Office with how long it has waited. Reading it does not clear it; you reply, approve or bench.
Running them Dispatches new sessions and runs them with no terminal open. Peek at one, reply from the list, or attach to the whole conversation. Reads the sessions your tools run and replies to them. It can start one. It never stops one, and closing it stops nothing.
Which tools Claude Code sessions. Its pages name no other tool. Claude Code, verified. Codex, verified for reading and replying. Gemini CLI and OpenCode adapters are written and unverified.
What you look at A list in your terminal, grouped by state, with the sessions that need you at the top. A research preview. An office seen from above in a window of its own: one room per repository, one robot per session.

As of October 2026. Agent view is described from Anthropic's own pages as they read on the 10th: the announcement and the documentation. DeckHQ is described from its README. deckhq doctor prints agent view's own count beside its own.

Studio turns an idea into an office.

Off until you turn it on for one project. Then a real session interviews you and writes the plan, the roster and the board into your own repository.

The Studio board: six columns named Backlog, Ready, In progress, Review, Done and Blocked. Each card has a title, the robot assigned to it, a count of criteria and its tokens. One card in Review is flagged waiting on you, handover.
The board on the demo floor: six columns, one card per piece of work.
  1. Describe the idea.

    deckhq studio enable <dir> --yes turns it on for one project. Then Studio: plan this project starts a real claude session that asks what you are building.

  2. Get a plan.

    It writes the blueprint, the roster and the board as files in that repository.

  3. Hire.

    A role gets a git worktree of its own, a brief that is yours to edit, and a real session there, on the floor within one scan.

  4. Run the board.

    A column is yours. Nothing DeckHQ observes moves a card; you do, by drag or with [ and ].

  5. Take the handover.

    A role that thinks a card is done writes a handover. The card is flagged, and its review shows the handover beside the worktree's diff. Accept it, or bounce it with a note.

  6. Keep track.

    Each card shows its tokens, its time in progress and the tests its handover quotes. A card past its budget goes to Blocked, and nothing is killed.

The loop is complete since 1.6.0. A real planner interview, and a hired session's real reply, have not been watched yet; Codex, Gemini CLI and OpenCode roles are marked unverified launch. Studio, in full

The details people notice.

Everything on the floor is a session that exists on your disk. Nothing is animated to look busy.

A crew: a lead robot named Elif at its desk, and five juniors sitting cross-legged in an arc behind it, each with a laptop and a cable running to the desk. A chip at the desk reads 2/5 working.

A lead with a crew.

A session that starts three or more sub-agents turns its desk into a formation: juniors cross-legged in an arc, a laptop each, and a cable to the desk that pulses while that junior is still writing. Twelve are drawn; beyond that a +N chip.

Two small robots, Yogesh and Hooman, sitting on the floor behind their lead, Boris, each with a laptop in front of it. Each name tag has two rows: a dark chip reading Junior, and the name under it.

Juniors say so.

A sub-agent is drawn one size smaller than the session that started it, sitting on the floor with a laptop, under a two-row tag: a Junior chip over its name.

Two benches against the wall of one project room. Kobe and Emeka sit side by side at a bench labelled fix/refund-rows; Ravi sits at a second bench labelled with another branch name.

A worktree is a bench, and the branch is written on it.

Five checkouts of one repository are one room. The main checkout has the desks in the middle; each git worktree somebody is working in is a bench against the wall. So is a repository mounted inside another one, with its folder's name on the bench.

The checkout-flow room with its lights off: a grey room with one empty desk. Its header counts 2 that need you and nobody working.

The lights go off.

When every session of a project is waiting in your office, its room keeps one empty desk and goes dark until one of them goes back to work.

Eleven styles, one click apart.

Look in the header repaints the whole floor: materials, furniture, partitions and light. Your agents keep their colours, because colour is state.

The floor in Studio oak, the default: warm herringbone wood in Your Office and the lounge, pale carpet in the project rooms.
Studio oak
The same floor in Night lab: dark floors and dark furniture, with the robots and their state colours standing out.
Night lab
The same floor in Colour plan: a neutral building where each project's room has a calm colour of its own.
Colour plan
The same floor in Graphite loft: grey terrazzo and loop pile, dark frames behind glass partitions, under evening light.
Graphite loft

All eleven, and the light

Before you install anything.

npx deckhq doctor

It counts the sessions on your disk. Before DeckHQ has ever run on a machine it writes nothing, starts no daemon and opens no window: it reads your transcripts and asks claude what is running and which MCP servers it has. Read the on the floor row: that is how many you have, finished ones included.

  • It listens on 127.0.0.1 and nothing else. There is no --host flag and there never will be one.
  • Zero dependencies. The package installs itself and nothing under it.
  • MIT, and all of it is on GitHub. No account, no analytics, no telemetry, no update check.

What it never does, and how to check

$ npx deckhq doctor
  claude code     available
  transcripts     77 sessions across 20 projects
  running now     6   (claude code's own agent view reports 6)
  on the floor    77  ← 71 sessions have already finished; the agent view no longer lists them
  codex           not installed
  waiting on you  0   (3 waiting, all still running)
  hooks           installed, port 4400, 285 events, last 1m ago
  state           ~/.deckhq/state.json, writable
  egress          none by design: only 127.0.0.1 (from the code, not measured on this run)

A real run, 3 September 2026. The report prints more rows today, and your numbers will differ.

One command.

It starts DeckHQ, opens the floor in a window of its own, and asks once whether you want a desktop icon.

npx deckhq app

Needs Node 18 or newer. Nothing else: the package has no dependencies.

No Node on the machine

One line checks for Node, offers to install it, installs DeckHQ and opens the window. Both scripts are short files you can read first.

irm https://deckhq.dev/install.ps1 | iex

curl -fsSL https://deckhq.dev/install.sh | sh

A step at a time

Install it, open it, then add the icon when you want one.

npm install -g deckhq

deckhq app

Look before you install

The report prints your own number. On a machine DeckHQ has never run on, it leaves no file behind.

npx deckhq doctor

Every way to install, and how to remove it