fleet
Create sandboxes on the owner's intentic account from inside this one, via the `sandboxes` CLI, read what the account already runs, and remove one a task is done with. Use when a task needs a SEPARATE
SKILL.md_sandbox/sandbox/seed-skills/fleet/SKILL.md
GitHubThe owner's other sandboxes
This sandbox can act on the account that owns it: list its sandboxes, create new ones on the owner's own
computers, and take one away. You reach that with the sandboxes command. The provisioning token behind it lives with the
daemon and is never in your environment, so nothing you can read or run creates or removes a sandbox without the
owner saying yes to that exact one, in chat, first.
Commands
sandboxes ls # every sandbox on the account, and whether it ever came up
sandboxes create storefront \
--on rog \ # which connected computer builds it, by its device name (the
# one its tools carry, mcp__rog__…); default: this one's
--definition ./team/reviewer.sandbox.toml \
--why "one line on why this job needs its own box"
sandboxes remove storefront --why "…" # take one off the account, the owner's card first
create holds its connection through both slow halves — the owner deciding, then ic pulling an image and
starting a container out on the machine. Minutes, either one. Hold it open rather than timing it out; every
line the machine printed comes back with the answer, and when a create fails those lines are the diagnosis.
When a new sandbox is the right answer
A sandbox is a machine with its own environment, its own connected services and its own workspace. Reach for one when the difference is any of those:
- Its own tools and access. A role that needs a database connector, a different toolchain, or credentials this box has no business holding.
- Its own workspace. A second project, whose repos and history should not be mixed with this one's.
- Its own blast radius. Work you want to be able to throw away, or that should not share this box's cores.
If the difference is only how a turn behaves — which model, which folder, which tone, which tools of the ones already here — that is a persona in this sandbox, not a new machine. Personas are free; a sandbox is a container somebody's computer has to run.
How consent works: enforced, not promised
Every create raises a card in the owner's chat naming the sandbox and the machine it would run on, and the call waits there. The order matters and is the point:
- The card comes first. Nothing is minted until it is answered yes. A no, or a card nobody answers, leaves the account exactly as it was — no row, no setup code, nothing to expire.
- Then the claim, spent at once. A yes mints the sandbox's row and a single-use setup code, and that code goes straight to the machine. It is short-lived; nothing holds one.
- Then the machine builds it, with
ic, under the same run contract every other sandbox on that computer uses.
So a refusal is an answer, not an obstacle: report it and carry on without the sandbox. Do not ask twice for the same one in a conversation.
What it will not do
- It never provisions hosted machines. A new sandbox runs on a computer the owner has connected as a device. Nothing here can spend their hosted plan or their money.
- It removes only behind the same card, and never for good.
sandboxes removenames the sandbox and what goes on the card. On a machine this sandbox reaches it runsic sandbox removethere, which keeps its files for a week; a sandbox that never came up, or whose machine is out of reach, leaves only the account, and Recently deleted brings it back for a week. It never removes the sandbox it runs in, a hosted one, or one of two that share a name (name it by its id). A sandbox you made outlives this conversation unless you clean up: when a task made one only to test something, offer to remove it at the end. - It carries no credentials into the new box. A definition names capabilities and secrets; it never carries their values, and the new sandbox's owner connects them there.
Definitions: a sandbox as a file
--definition takes a sandbox.toml — the same document Sandbox ▸ Environment exports. It names the
workspace repo, the repositories to clone, the capabilities the box should have, the secret NAMES to fill in,
the overlay Dockerfile, and the agent settings that differ from their defaults. The new sandbox applies it on
first boot, on a workspace that arrived empty.
That is what makes a team of specialized agents a reviewable artifact rather than a sequence of clicks: write the definition, commit it, and create from it. What a definition cannot carry is consent — the overlay lands there as a draft the owner still approves, and credentials land as names to fill.
Handing the new sandbox work
Creating a sandbox does not give this one a way to hand it tasks. That is a link, which the owner pairs in their
editor once the new box is up: its terms (the persona its tasks wear, whether each waits for its owner, its hourly
limit) are set over there. Then delegate reaches it; see the links skill. Ask for one with
capabilities request link --why "…".
If there is no account connection
sandboxes answers with a sentence saying so. Connecting it is the owner's, on a card:
capabilities request fleet --why "so I can bring up the reviewer sandbox we planned"
They mint a provisioning token at Settings ▸ Tokens and paste it into that card. Until they do, a second sandbox is something to ask for in words, not something to reach for.