TNTHUBTrigger & Trial
Close the laptop. Your agent keeps working.
TNTHUB gives you an isolated machine with SSH access. You connect and run whatever you want inside it — including an agent that has to keep working when you are not there. The same machine is where the specification pipeline will eventually run.
Free to use. No account required.
What the box is
A machine of yours, on our compute, with a shell. Nothing more is imposed on you.
A box comes up with an SSH server already running. You connect through the platform and you run your own commands — there is no runtime layer, no agent lifecycle manager, and no opinion about what the box is for. The default image has the usual build tooling in it; anything else you need, you install.
# Create a box
tnt box create --name research
# Open a shell in it
tnt box ssh-proxy ssh research
# Run one command and come back
tnt box exec research -- npm test
# Throw it away
tnt box destroy research
tnt box ssh-proxy ssh starts a local proxy that carries the connection and hands you an ordinary SSH session. Your editor, your terminal multiplexer, scp and your ~/.ssh/config all work exactly as they do against any other machine — and the box never needs an inbound port on the public internet.
Why it exists
The problem is small, concrete, and has annoyed every person who runs an agent on a laptop.
An agent stops when your laptop sleeps. You close the lid to go and do something else, and the work stops with it — not because the task failed, but because the machine underneath it went to sleep. You come back and start again, having lost whatever context was only in the process.
Put the agent in a box and the box does not care whether your laptop is open. That is the whole product today.
What people run in it
An agent that has to finish
Start your coding agent in the box, let it work through something long, and close the laptop. Come back later and read the log.
A long build
A test suite that takes forty minutes does not have to survive a dropped connection or a sleeping screen.
A service to point at
A staging API, a fixture server, a database at the exact version you need — reachable and disposable.
Something you would rather not run locally
A package that might misbehave, a script from a stranger, a reproduction you do not want on your own filesystem.
Running it here, versus running it there
| On your laptop | In a box | |
|---|---|---|
| Survives a closed lid | No | Yes |
| Survives a dropped connection | No | Yes |
| Reachable from elsewhere | No | Yes, over SSH |
| Cleanup | Whatever you remember to do | Destroyed, automatically when idle |
| Isolation | Your filesystem, your shell history | A machine of its own |
| The cost of a mistake | Yours to clean up | Gone with the box |
What comes after that
The box is the substrate. The specification pipeline is what we are building on top of it.
We are building a specification-driven delivery pipeline: the model drafts a specification, you approve it clause by clause, the platform builds from the approved snapshot, and verification runs as four mechanical layers. The build will run inside this same kind of machine, and you will still be able to get a shell and look at what happened.
It is a large piece of work and it is not finished. Here is exactly what exists today and what does not.
The method we build with
This part is not a promise. It is how we work, and you can read it now.
Prompting a model is a lottery: you cannot review what you cannot see. A specification is a contract. It can be diffed, approved, executed, and held against the implementation that was supposed to satisfy it. Our own development runs this way — the canonical specification is the source of truth, and the checks that guard it are executable.
Feature: Refund a settled order
Background:
Given an authenticated customer
And an order "A-1042" settled 3 days ago
Scenario: Full refund within the window
Given the refund window is 30 days
When the customer requests a refund for order "A-1042"
Then the refund amount is 42.00 EUR
And the order status becomes "refunded"
And a refund receipt is emailed to the customer
Scenario: Refund is refused after the window
Given the refund window is 7 days
When the customer requests a refund for order "A-1042"
Then the request is rejected with code "REFUND_WINDOW_CLOSED"
And the order status remains "settled"
The second scenario is the one that matters. A single happy-path assertion would have passed either way — see assertion strength.
Where this method fits
The platform will be strongest where behaviour can be stated precisely and checked mechanically. These are the domains we are tuning for first.
Payments & refunds
Money movement has crisp invariants. A wrong amount is a bug you find in production, not in a code review.
API contracts
An OpenAPI document is already a testable contract. Use it as one.
Risk & trading rules
Decision logic with numeric thresholds is the clearest possible specification.
Device fleets
Command handling, retry semantics and offline behaviour all have exact expected outcomes.
What it does not do
We would rather state the limits plainly than let you discover them later.
It is not a place to keep things
A box is disposable. Use it for work, take your results out, and destroy it. Nothing is inherited by the next person or the next tenant.
It does not manage your agent
There is no runtime layer. Your agent is a process you start; when it exits, it exits. The box gives you a shell and gets out of the way.
It is not a shared team environment
A box belongs to whoever created it. Sharing one means sharing credentials, which means losing the boundary that made it useful.
It does not take custody of your code
Your code goes in and comes out under your control, over your own SSH session. We keep the operational metadata needed to run the machine, and nothing more.
Under the hood
A dedicated machine
A box is not a shared container with your neighbours. Two boxes do not share a filesystem, and neither can see the other’s processes.
Reachable, not exposed
You connect out to the box; the box never listens on a public port. There is no inbound firewall rule to get wrong.
Idle reclamation
A box nobody touches for the idle window is torn down automatically, and its contents go with it. We publish the window in the open.
Snapshots
A reclaimed box can be re-provisioned from its snapshot, which is fast, because the snapshot and the machine image are immutable.
Frequently asked questions
Do I need to learn a new specification language?
No. The method uses Gherkin feature files and OpenAPI documents. If your team already writes BDD scenarios or keeps an OpenAPI document, you are already set up. If not, the template library gets you started in an afternoon.
What can I run in the box?
Whatever you want. The default image ships with the usual build tooling and a package manager; anything else you install yourself, and it persists for as long as the box lives.
What happens to my files when the box goes away?
They are gone. Destruction is not a soft delete and there is no recycle bin. Take out anything you want to keep before you destroy a box.
Does the specification pipeline exist yet?
No. The box is the first thing we are building and it is being built now. The pipeline is still being designed, and the status page says plainly which parts are which.
Can I use it with an existing codebase?
Yes — over SSH, with the tools you already use. And when the specification pipeline arrives, the useful unit is the specification, not the repository: most teams start by writing a spec for one vertical slice.
Start with one thing that has to keep running
Find the task you have been re-running because your machine went to sleep. Put it in a box.