Claude Code manager for Windows, macOS and Linux

Claude Code
instances manager
for your projects.

Claudeck starts, resumes, schedules and watches every Claude Code session on your PC. Each project keeps its own instances, each instance shows what it is doing, and you get a message on screen or on Telegram the moment one of them needs an answer. Idle instances go to sleep and give the memory back, your Claude accounts switch in one click, and the machine can turn itself off when the last job is done.

No card, no crippled features Windows · macOS · Linux 15 interface languages One or several Claude accounts

Go on, click it. Drag an instance onto the terminal to split the view. Right-click a row or a project, add an instance with the plus, wake a sleeping one, schedule a prompt, open the history, the settings, the statistics, the workspace menu with both accounts, the power menu, the bell, the notes. The prompt itself is the one thing this demo leaves to the real app.

145features, all of them listed below
15UI languages
1click to switch between your Claude accounts and workspaces
0lines of your code leave the machine
1

You are not at the keyboard

Claude works for minutes at a time, then stops on a yes/no question. You come back an hour later and find it waited 58 of those minutes for one keystroke.

2

Idle sessions still cost you

Every parked Claude Code process holds hundreds of megabytes just by existing. Ten of them and your machine is paging while nothing is happening.

3

Limits arrive without warning

You fire off the big refactor with no idea that 92 % of the weekly Opus window is already gone, and find out when the run dies half-finished.

How it works

Start the work, leave the desk,
come back to a machine that switched itself off

Claudeck plugs into Claude Code's own hook stream (SessionStart, UserPromptSubmit, PreToolUse, Notification, Stop) plus the status line. It is not watching pixels or guessing from output. It knows.

STEP 1 You One prompt, then you walk away. refactor the API layer STEP 2 Claudeck holds every session and reads Claude Code's hooks migrate the hook server Working… 4:12 landing page copy Working… 1:08 database migrations Waiting for you old spike branch Asleep · 0 MB z z z RAM HELD BY IDLE SESSIONS 3.1 GB 0.8 GB after sleep STEP 3 It reaches you wherever you are Desktop card, sound, and Telegram on your phone. 11:40 1 Claudeck · Waiting 11:40 database migrations · acme-api Should I drop the legacy usage_daily table? STEP 4 When the last job is done, so is the PC Background subagents counted in. 60-second countdown with Cancel. Sleep the PC: armed Working instances right now: 2 Telegram queue flushed first

Swipe sideways to see the whole diagram

Working Waiting for you Ready Asleep: no process, no RAM Ended
The parts that change your day

Twelve things you will use every single day

Everything below is the real interface, rebuilt in this page from the app's own markup and stylesheet, not a screenshot. Change the look in the top-right corner and it re-tints, exactly like the app does.

Every session in one deck, and you can read it in one second

Projects on the left, their instances underneath, the live terminal on the right. Each instance is a real claude process with its own terminal, created the moment the instance exists: switching sessions only toggles visibility, so no output is ever lost and nothing re-renders.

Status you can trust. The second line is driven by Claude Code's own hooks plus the status line, not by scraping the screen. Working shows how long the phase has lasted; a question turns the whole row amber so you cannot miss it.

Sessions survive everything. Every instance gets a pre-generated --session-id, so a reboot, a crash or an update brings all of them back with --resume, exactly as they were. An instance that exits shows a bar with the exit code and a Restart button.

They name themselves. After the first prompt an instance adopts the title Claude generated for the session: typed into the panel letter by letter. Rename it yourself and it never changes again.

Nothing is ever thrown away. Closed instances go to a searchable list with cost, tokens, messages and the first prompt, and Restore resumes the very same session.

Several instances in view at once

Put instances side by side or one above the other: right-click one and pick a split, or drag it from the sidebar onto the terminal. The half it will take lights up before you let go.

Every new pane takes its room from the others in proportion, so a 40 / 60 layout keeps its ratio.

Drag a divider to resize, drag a header to move a pane. The cross closes the pane and the instance keeps running.

Every pane counts as open: a sleeping instance wakes up when it joins, and nothing in view is put to sleep.

The prompt tools act on the pane you clicked, and every workspace keeps its own layout across restarts.

An idle session should not cost you a gigabyte

An instance that is not open, not working and untouched for the time you choose, or the one that has been idle the longest once all your instances together exceed the memory limit you set, loses its Claude process and stays in the deck as Asleep. Open it and Claudeck restarts Claude in the background, resuming the same session, and reveals the console once it is ready. You lose nothing but the memory.

zzz
The instance was put to sleep
Claude was shut down to save memory because the instance had not been opened for more than 15 minutes and was not working.
Starting it again in the background…
Sleeping
Over a memory limit
Memory limit
4.0 GB
Sleep working instances too

Three modes. After idle time: an instance that is not open, not working and unchanged for longer than the chosen time (5 minutes to 12 hours) is shut down in the background. Over a memory limit: every 15 seconds Claudeck adds up the memory of every instance's processes; over the limit the instance that finished working the longest ago is put to sleep, the sum is measured again, and so on. The open instance never sleeps; working ones only with the switch. Off for a PC with memory to spare.

Manual too. Right-click → Sleep parks the session you will need tomorrow and reclaims the RAM now. Sleeping instances stay asleep across an app restart, and Restart in the same menu gives a stuck one a clean process without losing the session. Never sleep, a switch in the same menu, keeps one instance awake in every mode: over the memory limit the next one in line sleeps instead.

A budget instead of a timer. Pick Over a memory limit and drag the slider from 0.5 GB up to 80 % of your RAM. Claudeck measures the real process trees (Claude Code, its MCP servers, the Node helpers), also right after every start, and puts instances to sleep least-recently-working first until the sum fits. The one you are looking at never sleeps.

No zombies left behind. Closing or sleeping takes a WMI snapshot of the whole process tree and kills every surviving PID, including the npx-bootstrapped MCP servers that normally pile up in Task Manager.

It finds you. You stop checking.

Mark an instance with the bell and Claudeck messages you the moment it finishes or starts waiting for you: through one company bot, so there is no BotFather, no token to paste and no webhook to host. Thirty seconds of setup: press Start once in Telegram.

Claudeck · Done11:42

migrate the hook server · acme-api
Every hook request now carries the per-run token and the old route answers 404. All 94 tests pass.
Claudeck · Question11:40

database migrations · acme-api
Should I drop the legacy usage_daily table, or keep it and add a view?
Claudeck · Error11:31

flaky e2e run · acme-api
The work did not finish, the weekly limit is spent. It comes back Mon 09:00.

The desktop card waits until you are actually back: its 8-second countdown only starts after your first keypress or mouse move on the computer. A card that arrives while you are away is still there when you return. Six kinds (Done, Question, Plan approval, Permission needed, Error, Needs attention) each from a real Claude Code hook.

Send for
Marked instances only (the bell under the terminal)

Notify on Telegram

When the work ends Waiting for you

Written for a lock screen. Emoji, bold instance name, project, then one plain sentence in your UI language. No tool names, no error codes. Questions quote Claude directly, and a Done message says how long the run took.

Start the long job at 18:00. Let the machine turn itself off.

Arm Sleep, Hibernate or Shut down from the power button in the title bar. Once every instance has stopped working (background subagents and background shell commands included) Claudeck waits 30 seconds of quiet, shows a one-minute countdown with Cancel, flushes the queued Telegram messages, and then does it.

Nothing is working yet: the options stay out of reach, so an idle desktop can never be shut down by accident.

After all work is done

Once every instance has stopped working, including agents running in the background, Claudeck performs the chosen action. A one-minute countdown with a Cancel button runs first.

Working instances right now: 0

To activate this feature an instance must be working.

Two instances working: now you can arm it, and a green dot on the power button says it is armed.

After all work is done

Once every instance has stopped working, including agents running in the background, Claudeck performs the chosen action. A one-minute countdown with a Cancel button runs first.

Working instances right now: 2

The last one just finished: thirty seconds of quiet, then this card. Cancel disarms it, Do it now skips the wait.

All work is done

The PC goes to sleep in 47 s.

The rule that makes it trustworthy: Loading and Working count as busy, Waiting for you does not. A forgotten question cannot hold your machine hostage all night, and a Stop held back for background tasks keeps the instance in Working, so real work is never cut off mid-flight.

Two Claude accounts on one PC. One click to switch.

Sign in with as many Claude accounts as you have and pick one per workspace from the menu in the title bar: the workspace's instances restart on the other login and resume the very same sessions. A workspace owns its own projects and instances; the others keep running in the background like windows on another virtual desktop. An account is just a Claude Code config directory handed to its processes in CLAUDE_CONFIG_DIR, so both are logged in and working side by side. Try it below: click a workspace or an account.

Workspaces

Own projects and instances, each workspace on one Claude account.

Claude account

The workspace “Workspace #1” runs on this account

The weekly limit is gone? Move the workspace. Picking the other account restarts that workspace's instances and resumes the same sessions, on the other subscription. The one you were on stays as the default for next time, or pin one for good.

Sign in inside the app. Adding an account runs claude auth login in a real terminal inside a dialog; finish in the browser or paste the code. Claudeck then reads the e-mail, organisation and plan.

Same setup on every account. A new account folder gets directory junctions to your projects, plugins, skills, commands, agents and hooks, and a copy of your settings and CLAUDE.md. Only the login differs.

Per-instance Claude flags

34 catalogued claude options as form fields, plus a free-text box and a live preview of the exact command.

claude --model opus --effort high --dangerously-skip-permissions --resume 6f1c…

Tell Claude what to do at six, or the moment it finishes this one

Right-click an instance, Schedule a prompt: pick the day in the calendar, spin the time, or choose After the current work while the instance is still busy. The prompt is pasted and sent only to a Ready instance: a sleeping one is woken first, a working one gets it after its turn, and a moment that passed while Claudeck was closed is sent at the next start. Up to 50 prompts per instance, each editable or cancellable from the list. Try it: the form below is live.

An alarm clock on the row. Every instance with something waiting shows a small alarm-clock mark; hovering it tells you how many prompts wait and when the next one fires, and the context menu carries a red badge with the count.

Only ever to a Ready instance. Asleep? It wakes up first. Working? It gets the prompt right after. PC off at the time? The prompt goes out at the next start. Multi-line prompts travel through bracketed paste, so they never leave line by line.

Fast to fill in. Chips for +15 min, +30 min, +1 h and Tomorrow 8:00; hours spin by one and minutes by five with the wheel or the arrows; a green line says Sends in 2 hr 15 min, a red one says the moment has passed. Ctrl+Enter schedules.

Which instance did I write that in? Every prompt, with what Claude did next

Right-click an instance, Prompt history: Claudeck reads the session transcript and lists every prompt newest first, grouped by day, with the time, the tools Claude used, the files it touched and its last words. A card opens to the file list, Copy and Put into the prompt. Search across the prompts, narrow the list to one day with the picker or the calendar, and read the session like a diary: messages typed while Claude was working, slash commands with their output, and the moments the instance slept and woke up.

Nudges you sent mid-run. A message typed while Claude worked, which Claude Code keeps only as an attachment to the next prompt, shows as its own dashed card labelled Added while working, with an arrow to the prompt it joined.

Commands included, not counted. /model, /compact or !ls appear as quiet grey rows with their output, searchable but outside the prompt count.

Big transcripts, no wait. Transcripts run to megabytes; Claudeck streams them and shows a loader while it reads, then a day picker with counts and a calendar with dots on the busy days.

Know what is left before you fire off the big job

One full-width bar per limit your account actually has: the 5-hour session window, the weekly all-models window and each model's own weekly window. The list comes straight from the usage endpoint, so a new model or a new kind of limit simply appears. Orange from 75 %, red from 90 %.

Session1 hr 32 min
43 %
All models4d 0 hr
78 %
Fable3d 22 hr
92 %
Sonnet6d 3 hr
41 %

Two independent sources. The same endpoint /usage uses, polled every minute, plus a status-line relay as a backup, and your own status line (ccstatusline and friends) still renders exactly as before.

A spent limit is not success. When Claude Code dies on a quota error, Claudeck says the work did not finish and tells you when the window comes back, instead of letting you assume it shipped.

History per account. Every change of every limit is recorded and drawn over 12 hours, a day, a week, a month or a calendar month. With two subscriptions you can finally compare them and route the work to whichever has headroom.

Only the bars you want. A checkbox per limit in Settings hides the ones you never look at; a limit that appears later is on by default.

up to 75 %75–90 %over 90 %flat line = no change or Claudeck off · vertical drop = window reset

Session

now 43 % · reset in 1 hr 32 min

How much work you actually got out of Claude Code

Every interval an instance spends in Working is logged with a 30-second heartbeat, so the numbers survive a crash. Parallel time is counted honestly: three instances working 8, 10 and 12 minutes at once is 12 minutes of work, of which 10 were parallel: a day never becomes 30 hours.

Statistics

Last 30 days
Today
3 hr 32 min
of which in parallel 1 hr 4 min
This week
18 hr 4 min
of which in parallel 4 hr 51 min
This month
72 hr 51 min
of which in parallel 19 hr 20 min
Average per day
4 hr 6 min
last 30 days
Streak
23 days
days with work in a row

Trend

last 30 days · lighter = work, darker = of which in parallel

Activity

last 52 weeks · at most 6 hr 12 min per day
Less More

Projects

last 30 days · click to open the detail

Billable to the interval. Click a project and drill down: days, then instances, then single intervals (“23 min · 10:30–10:53”) with a month picker or a custom range.

One keystroke away. Ctrl+Shift+S opens the page over the terminals, Ctrl+Shift+U opens it on the Usage tab, Esc closes it. The instances keep running underneath.

Say the prompt. Nothing leaves the machine.

The microphone records 16 kHz mono, and whisper.cpp transcribes it locally: no system dialog, no cloud, no API key. Audio is cut into chunks that are transcribed while you are still talking, so a two-minute dictation does not mean a two-minute wait: after you stop, only the last ~2 seconds are awaited.

Speech to text with Whisper

Claudeck turns your voice into text with Whisper, the most accurate speech recognition there is. It understands more than 100 languages, runs entirely on this computer, and nothing you say leaves it.

Whisper is not part of the installation, because not everyone needs it: it is a one-time download of about 547 MB. Download it now?

Chunked, gapless. A chunk ends at the first 300 ms pause after 12 s (forced cut at 30 s); the next one starts a second earlier, the overlapping words are removed, and the tail of the previous transcript is passed in as context so sentences and punctuation carry across.

Verified on real speech. 66 seconds of dictation: word-for-word identical to transcribing the whole file at once. The model (Large v3 Turbo) lives outside the app folder, so updates never re-download it.

24 colours, 9 fonts, 16 shortcuts, 15 languages

Every colour in the app is a variable you can change, with a native colour picker and a hex field. Buttons, hovers and borders are derived from the panel colour until you override them, so one colour usually retints the whole app. Try it: the four looks below drive this whole page through the very same tokens.

Settings

Appearance

Font and colours of the whole app. Changes apply at once and save themselves.

Font
Nunito

Colours

Look
Graphite

Pick a look and fine-tune the colours. Click a swatch to open the colour picker, type a hex next to it; the arrow next to a changed colour restores the template value. An adjusted look can be saved as your own template.

Panels

Left panelButtons, highlights and borders are derived from it until you set them.
Right panel and terminal
Title bar
AccentWorking, the active switch, the primary button.

Elements

Buttons
Hover highlight
Active instance and selection
Borders and dividers
Ready
Asleep

Keyboard shortcuts · projects and instances

New projectCtrl+N
New instance in the active projectCtrl+Shift+I
Next instanceCtrl+Tab
Rich input for the promptCtrl+Shift+P
Dictate into the prompt (start / stop)Ctrl+Shift+M

Shortcuts you can actually remap. Sixteen in three groups: click Change, press the combination, Esc cancels. Combinations are read from the physical key code, so a Czech or German keyboard gives the same letters as a US one, which most shortcut systems quietly get wrong.

15 languages, completely. English, Czech, German, French, Spanish, Brazilian Portuguese, Italian, Polish, Ukrainian, Turkish, Indonesian, Vietnamese, Japanese, Korean and Traditional Chinese: tray menu, notification window, hook texts and Telegram messages included.

Nine fonts, bundled where needed. Nunito, Open Sans and Roboto ship with the app, plus Segoe UI, Segoe UI Variable Display, Calibri, Verdana, Trebuchet MS and Consolas. The terminal stays monospace.

Time and resources

The savings are not a slogan. They are arithmetic.

Two numbers decide how much Claudeck gives back: how many sessions you keep open, and how much of the day they sit there doing nothing. Move the sliders.

6
5 h
3

Assumptions, all conservative: ~420 MB of resident memory per idle Claude Code process (Claudeck frees all of it when an instance sleeps), a desktop drawing ~75 W while it sits there overnight, 8 unattended hours per evening, and roughly 30 % of the idle-but-open window being a session that was already finished or already blocked on a question.

RAM handed back
2.1 GB
while those sessions are parked, and nothing is lost, they resume where they stopped
Dead waiting removed
1 h 30 m / day
about 31 h a month: time a session sat finished, or blocked on a question, before you noticed
Power not burned overnight
7.2 kWh / month
because the machine turned itself off when the last job finished, not when you woke up

Tokens you do not pay twice for

A reboot, a crash or an app update resumes every session with --resume. No re-explaining context, no re-reading the same files, no rebuilding a conversation you already paid for.

No premature “done”

Claude Code fires Stop the moment the main agent answers, even while background subagents are still running. Claudeck reads the background_tasks list and holds the “done” until it is empty.

No RAM leaked to zombies

Closing or sleeping kills the whole process tree, MCP servers included. After a crash, orphaned Claude processes from the previous run are found and cleaned up before the sessions restart.


The complete list

145 features. Tap any of them.

The whole product, grouped the way the app is built, nothing summarised away. Pick a chip, search, or let it surprise you: the card on the right tells you what it does and why it is there.

Seen 0 of 130

Nothing matches that. Try a shorter word.


Under the hood

It sits on top of your source tree.
It behaves accordingly.

Hooks on loopback only

Hook and status-line events arrive on a local HTTP server bound to 127.0.0.1 on a free port, with a per-run secret token in the path. Anything else gets a 404, bodies are capped at 1 MB, and Claude is answered before the payload is processed, so it never waits on Claudeck.

Nothing in the UI phones home

The renderer declares default-src 'self': no remote code, no remote assets, images and fonts limited to self and data URIs. Whisper dictation runs locally. Your prompts and your code go to Claude, and nowhere else.

Licence checks that tolerate reality

The licence is an Ed25519-signed token verified offline against a key baked into the app. It refreshes after three days and carries a 14-day deadline, so you can work on a plane or behind a hostile proxy for two weeks without the app locking.

No JavaScript error box, ever

Uncaught main-process errors, renderer errors and rejected promises are appended to a rotating log under the app data folder and the app carries on. A bug in the shell never interrupts a Claude session that is mid-task.

State written on every change

Projects, instances, session ids, order, notes, settings and the window position are saved the moment they change, never only at quit. A corrupt state file is backed up with a timestamp and replaced rather than blocking startup.

Updates you decide on

An orange dot on the gear, a bar in Settings, a “what is new” link and a Download button. The installer is verified by sha256 and size before it runs. Nothing is ever downloaded or installed without a click.

Pricing

A month to decide. Then one key per computer.

Claudeck is the shell. You keep using your own Claude subscription: nothing is resold, no tokens pass through us.

Free trial

30days

No card, no sign-up, nothing disabled.

  • The complete app: every feature on this page
  • Starts on first launch, no account needed
  • A slim countdown bar in the last three days: no surprise lock-out
Per computer

Yearly licence

€79/ year

One key = one computer. Move it whenever you like.

  • All updates for the whole year
  • Works offline for two weeks at a stretch
  • Deactivate on this computer frees the seat for a new laptop: one click, no support ticket
  • Your machine is identified by a salted hash: the raw machine id never leaves it
Buy a licence

Key format CLDK-XXXX-XXXX-XXXX-XXXX

Downloads

One Claudeck for Windows, macOS and Linux.

Pick your system. Same app, same licence key, same 30-day trial on every platform.

Every release carries its SHA-256 checksum on the changelog. The trial starts at first launch: no card, no sign-up.Release notes

Questions people actually ask

Do I need a Claude subscription?
Yes. Claudeck runs the official claude CLI on your machine with your own login: it is a shell around Claude Code, not a reseller. Nothing you send passes through our servers.
Does it change how Claude Code behaves?
No. It adds hooks through --settings so it is told when Claude starts, finishes, asks or needs a permission, and it relays the status line before running your status line command, so ccstatusline and friends look exactly as before. Claude Code still draws its own UI in its own terminal.
Does it resume a session after a usage limit?
No, and it says so plainly. When Claude Code dies on a quota error Claudeck reports that the work did not finish and, when it knows from the limit bars, when the window comes back. What it does resume automatically is a session after a restart, a crash, a quit or an instance sleep.
Is sleeping an instance the same as closing it?
No. Sleeping ends the Claude process and frees its memory, but the instance stays in the deck and its session id is kept. Opening it restarts Claude in the background with --resume on the same session. Closing puts it in the recently-closed list, from which Restore also resumes the same session.
Can it shut the machine down while something is still running?
It is built specifically not to. Loading and Working count as busy, background subagents and background shell commands keep the instance in Working, and after the last one stops Claudeck still waits 30 seconds of quiet plus a one-minute countdown with Cancel. “Waiting for you” deliberately does not count as busy, so a forgotten question cannot keep your PC on all night.
Two Claude accounts: is that allowed?
Each account is a normal Claude Code login in its own config directory, used by its own processes. Claudeck does not share or pool anything between them; whether you may hold two subscriptions is between you and Anthropic's terms.
Which systems does it run on?
Windows, macOS and Linux. The Windows build is out today; the macOS and Linux builds follow, and the download section shows each one the moment it is released. One licence key works on any of them.
Where does my data live?
State, per-account usage history, the work log and the licence are in the roaming app data folder, so updates and reinstalls never touch them. Claude Code keeps its own session transcripts exactly where it always does.

Run the deck for a month.

If you are not at the keyboard while Claude works, Claudeck pays for itself in the first week you leave the office with a job still running.