Start
$39a month
One long job all night, 3 terminals at once
- vCPU
- 2
- RAM
- 4 GB
- disk
- 50 GB
- Internet
- 50 Mbit
- Terminals at once
- 3
Close the laptop. Your agents keep working.
Hand Claude Code, Codex or OpenCode the long job and wake up to it finished. They come set up in the cloud on the Claude or ChatGPT subscription you already pay for, or your own API key, move to your next subscription when a limit runs out, and reach you in any browser, on your phone, in Telegram or by voice in ChatGPT. From $39 a month.
Get my Compute
Your subscription or your key. Claude, ChatGPT, or a key from one of seven providers. You pay us for the compute only.
Your first overnight run works, or your money back · Docker and sudo included
5000+ developers have moved their agent work off the laptop.
Ready at login
- Claude Code
- Codex
- OpenCode
- Docker + compose
- Ubuntu 26.04 LTS
Your subscription or your key
A limit runs out at 3 a.m. Your agent keeps going
Every 5-hour and weekly window sits in the header of every page with the minute it refills, so you know which subscription has room before the job starts, and the agent moves there when a limit runs out. Sign in with your Claude or ChatGPT subscription, or paste a key from Anthropic, OpenAI, xAI, DeepSeek, Z.ai, abliteration.ai or Gemini.
Seven providers by key
Paste a key and each launch offers the agents that take it. OpenCode takes all seven; Claude Code and Codex take the ones built for them.
Switch at the limit
Near the end of a limit, the agent moves to your next subscription and carries on: down your list in your order, or soonest reset first, so the usage about to expire gets spent first. A dry API key hands over to your next key from that provider. On a single subscription, it types “continue” the minute the limit resets. You wake up to finished work.
Keep a reserve
Keep 20% of a subscription's week for yourself: your agents move off it before they touch that 20%. On the last evening before the reset, lift the reserve with one press and let them spend the rest.
On your laptop, the agent dies with the lid
Close the lid and the refactor stops halfway. The limit you pay for resets at 3 a.m. and sits unused until you wake up.
A bare VPS keeps the agent running, but setup eats an evening: Docker, tmux, three CLIs, logins and a token file you copy by hand. And your phone still has no terminal you can work in.
What a night run needsVendor docs, read 17 Oct 2026
| What you get | detachbox | Who else has it |
|---|---|---|
| Moves to your next subscription when a limit runs out | Yes | None of the eight alternatives. In Orca you swap accounts by hand, for new sessions only. |
| Every subscription's limits on one screen | Yes | Only Orca, while it runs on your laptop or a server you keep on. |
| Picks up at the reset: Claude Code, Codex and OpenCode | Yes | Claude Code's own CLI, for Claude only, after resets under 24 hours. |
| Keeps working with no time cap | Yes | A bare VPS, Terminalbox and exe.dev, with no rotation and no limits screen. Codespaces stops at 12 hours. |
| A real terminal on your phone, laptop off | Yes | exe.dev, on iPhone only. Terminalbox's browser terminal has no key bar; Orca's app needs its desktop on. |
| Answer the agent from the notification, Telegram too | Yes | None of the eight says so in its docs, and none has Telegram. |
| Private HTTPS previews, public links for clients | Yes | Codespaces also offers private or public ports. exe.dev has a private proxy your client signs up for, plus one public port. |
Projects
Five agents at once, each in its own repo
Every repo is a project on your Compute: a folder the agent starts in, with your CLAUDE.md, AGENTS.md, skills and hooks, on the subscription or key you pick. A new project costs $0. Your size sets how many terminals run at once: open five on Pro and five agents work side by side.
Every agent on its own account
Claude Code on your Claude Max plan writes the API while Codex on a second ChatGPT account reviews it, on the same Compute.
A new project costs $0
Clone a repo, start empty or take a folder already on your Compute. One disk and one Compute plan cover every project.
Logins stay in their project
Each project has its own home folder, so its logins, configs and agent settings stay inside it. A project terminal sees only its own code.
Browser terminal
Start it at the desk. Pick it up anywhere
Close the laptop at your desk, open your phone on the train: the agent is still at work, and every line it wrote waits in the scrollback. Every terminal is a tmux session on your Compute, 3 to 30 at once by size.
Paste a screenshot
⌘V a screenshot and the agent gets it as a file. Drop other files onto the terminal, up to 10 GB each on Business Ultra, and the agent gets their paths. A dropped connection resumes the upload where it stopped.
Keys you already know
⌘C, ⌘V, ⌘K, ⌘F and Option as Meta, like the Terminal on your Mac; Ctrl+Shift+C, V and F on Windows and Linux. When the agent copies text, it lands in your clipboard.
Click a path, read the file
File names in the output open in a viewer with highlighting, ready to download, copy the path or share by link.
Terminals and notifications
The agent that needs you finds you
Every agent of every project in one list, each with its state: working, needs you, done, or stopped on a limit with the minute it resets. Your phone buzzes the moment one asks, and you answer from the notification.
Answer from the lock screen
The agent's menu arrives as buttons in a browser push or a Telegram message. Tap your answer and the agent carries on.
Know the moment a run ends
A ping when a run finishes or stops on a limit. An email follows if a ping sits unseen for ten minutes.
Quiet hours
Mute routine pings while you sleep. The out-of-memory email stays enabled, and quiet hours delay its delivery until they end.
Assistants
Drive your agents by voice from the car
Connect ChatGPT, Claude, Grok Bot or Meta Muse once. Ask what every terminal is doing and answer the agent that waits. Hand the next task to the one that finished, out loud on your phone or in any chat.
ChatGPT keeps watch
ChatGPT's always-on agent, a “dot”, writes to you the moment one of your agents asks or finishes. Call it back and answer by voice; whatever it starts on your word keeps running after you hang up.
Everything you do on the Terminals page
The assistant starts a terminal with a task, restarts an agent, switches it to another subscription and reads out the plan the agent just wrote.
One address, one click to revoke
One MCP address connects every assistant. Approve access or add an MCP key, and revoke either in one click. Billing and your sign-in stay out of their reach.
Previews and files
The agent's work on your phone and your client's
The agent starts a dev server, and the port already has its own HTTPS address on detachbox.dev, hot reload included. It opens for you alone until you share it.
Share one port
Make a port public and send the link. Make it private again and the link closes at once.
Reports as web pages
A coverage report, a built site or any HTML file on your Compute opens in your browser at its own link.
Files by link
Send one file to someone with no account: the link lives from 15 minutes to 7 days, and you see how many times it was opened.
Vault · powered by Sallyport Cloud
Keys your agents use and never hold
Your agent calls Stripe or GitHub, and the service key never enters your Compute: every project gets its own Sallyport agent, with only the access you give it in Sallyport Cloud.
Access per project
A new project starts with access to no service. The agents shipping your API reach Stripe; the ones on your docs site reach nothing.
Calls wait for your yes
A call that needs your approval waits in Sallyport until you give it; Sallyport asks you directly.
A token you can take back
The token lives in its project and works only from detachbox's addresses. Delete the project or disconnect, and the token stops working.
Ready at login
Ubuntu 26.04 LTS with sudo, set up the way you would do it, without losing the evening.
- Agents
- Claude Code · Codex · OpenCode · the latest versions, approvals off
- Containers
- Docker Engine with compose, you in the docker group
- Languages
- Node 26 · Python 3.14 · Bun · uv · mise
- Tools
- git · gh · glab · ripgrep · fd · fzf · jq · tmux
- Databases
- psql · redis-cli · sqlite3 · the MariaDB client
- Testing
- Playwright with headless Chromium · httpie · yq · bat · delta · git-lfs
- Access
- sudo · the browser terminal · git keys in each project's home
- Size
- 2 to 30 vCPU · 4 to 192 GB RAM · 50 to 1,000 GB disk
And the small things you stop doing by hand
- ⌘K or Ctrl+K to any terminal, project or account
- Text the agent copies lands in your clipboard
- Links the agent opens land in your browser
- Every package cache cleared in one click
- Waiting agents counted in the tab title
- A Codex limit reset spent with one button, when you choose
- An out-of-memory stop explained, with a one-click restart
- Every assistant action signed “You via ChatGPT” in Activity
Wake up to a finished refactor
A refactor overnight, agents side by side, a backlog turned into branches, client work, long data jobs. Pick the one that looks like your week.
All fourteen use cases and stories
Agents get sudo. Your service keys stay out of reach
The agent gets sudo inside its own terminal. Refresh tokens stay on our server, service keys stay in Sallyport, and each project keeps its logins to itself.
- Tokens
Refresh tokens never leave our server
Your Claude and ChatGPT refresh tokens stay encrypted on our server. Your Compute gets an access token we renew on schedule, so an agent with sudo holds nothing that can mint a new one.
- Projects
One project's logins never reach another
Every project has its own home folder. Logins, configs, agent hooks and caches of one project are invisible to the others, and a project terminal mounts only its own code.
- Vault
Service keys stay in Sallyport
Agents call Stripe, GitHub and the rest through the project's Sallyport agent. The key itself stays with Sallyport, and a new project starts with access to nothing.
- Sign-in
A copied session opens nothing
Sign in with a passkey, Google, or a password plus an authenticator code. Every dashboard request is signed by a key your browser made and cannot export, so a stolen session token works nowhere else.
- Previews
Previews private until you share them
Dev servers answer on detachbox.dev, a domain apart from your dashboard. Only you open them until you make a port public.
- Bet
Our bet: one kernel, many customers
Every terminal is an unprivileged system container on a shared kernel, with nesting on so Docker runs inside. Private networks, our servers, other customers and port 25 are blocked, so every customer is sealed off. The shared kernel is our bet, and it gets you full Docker at $39.
- Read the security model
Your agent works around the clock, from $39 a month
Bring your subscription or your key, we bring the compute. One Compute runs every project you have, around the clock: no hourly billing, no sleep timer, and the price on the card is the whole bill.
Pro
Our pick for agent work$75a month
5 agents plus the Docker services they test against
- vCPU
- 4
- RAM
- 8 GB
- disk
- 100 GB
- Internet
- 100 Mbit
- Terminals at once
- 5
Max
$169a month
8 agents at once and heavy builds on 8 cores
- vCPU
- 8
- RAM
- 16 GB
- disk
- 200 GB
- Internet
- 200 Mbit
- Terminals at once
- 8
Your first overnight run works, or your money back. If the first run on your first Compute doesn't work for you, write to us within 14 days of its first payment and get that payment back in full.
Charged when your Compute starts, then monthly in USD. A bigger size lands on the fly, no restart. Cancel from the dashboard and it runs to the end of the paid month.
Left on 24×7, a 2-core codespace with 50 GB comes to $134.90 a month at GitHub's rates. Start is $39 with Docker, every project you have and three terminals that outlive your laptop.
Running agents for a team?
Business adds more memory, more terminals at once, support tickets that go first in the queue and invoices made out to your company. Enterprise adds an SLA and a dedicated node we set up, or Enterprise on-premise, with every engineer's Compute on your own servers.
Questions before you start
What happens when I close the tab?
Nothing stops. The tab was a window onto a tmux session in your terminal. Open the dashboard from any browser or your phone and you land in the same session, scrollback intact.
Do I need an API key?
No. Claude Code and Codex run on the Claude or ChatGPT subscription you already pay for. Have keys? OpenCode takes them from all seven providers, Anthropic to Gemini, and Claude Code and Codex take the compatible ones. Connect them in the dashboard and every agent you launch starts signed in.
What happens when my subscription hits its limit?
The work carries on. With rotation on, the agent moves to your next subscription, in your order or soonest reset first; with continue on, it picks up the minute the limit resets, weekly limits included, in Claude Code, Codex and OpenCode alike. With both off, the agent pauses and your phone gets a ping.
Can I run several agents at once?
Yes, on one Compute. Each terminal runs Claude Code, Codex or OpenCode in its own project, on its own subscription or key, with that project's CLAUDE.md, AGENTS.md, skills and hooks. Your size sets how many run at once, from 3 on Start to 30 on Business Ultra. Need more at once? Change size, no restart.
Is it safe to connect my Claude or Codex subscription?
Yes. Your refresh token stays encrypted on our server and never reaches your Compute; the CLI gets an access token. Disconnect the account and we delete its tokens from our server and from your terminals.
Can I run Docker and databases?
Yes. Docker Engine with compose and sudo come in every terminal, and you are in the docker group. Containers run until their terminal ends or restarts; keep database files in your project folder and they outlast it.
Can I drive it from my phone?
Yes. The dashboard works in your phone's browser, with a key bar for Esc, Tab, arrows and quick answers. It has a photo key and voice input for prompts too. Add it to the home screen and it opens like an app, or connect ChatGPT once and talk to your agents out loud.
What if the server reboots?
Your terminals come back by themselves, with their packages, Docker images and files in place. Running processes stop, and each agent comes back on its last conversation by itself, waiting for your next word.
Why is there no free plan?
To keep the CPU you pay for working for you: crypto miners fill free tiers first. A friend's invitation makes your first month free; otherwise start at $39 and cancel from the dashboard whenever you like, and your Compute runs to the end of the month.
What if it doesn't work for me?
Your first overnight run works, or your money back. If the first run on your first Compute doesn't work for you, write to us within 14 days of its first payment and get that payment back in full.
Where does it run?
On servers around the world, and we connect more regions on request. The agent works next to its code, so distance only touches your keystrokes. Anyone, anywhere can sign up.
Can my team share one Compute?
A Compute belongs to one person. Each teammate signs up for their own; at $39 it costs less than the argument over whose agent is eating the CPU.
Can I get more than one Compute?
One per account, and it runs every project you have, so you never pick where code lives. Change size for more power, up to 30 vCPU and 192 GB on Business Ultra, without a restart. A whole team on its own node is Enterprise.
How do I cancel?
In the dashboard, through the billing portal. Your Compute keeps running to the end of the month you paid for. Free days and a free month end the moment you cancel.
Ship twice as much without working twice as long.
Hand Claude Code, Codex or OpenCode the long job and close the laptop. Check in from your phone, Telegram or ChatGPT while it runs on the subscriptions you already pay for.
Get my Compute$39 a month · Your first overnight run works, or your money back