Give your agent a task.
Describe a website or import a repository. Set the scope and let your agent get to work.
Your agents, at work. Every change, ready to review.
A workspace for running tasks, inspecting code, and keeping the whole
story.
From the first failing test to the final patch.
From an idea to a reviewed resultYOUR NEXT IDEA STARTS HERE
An idea. A task. A way forward.
THE WORKFLOW / 01—03Watch the work take shape.
Code runs in an isolated cloud sandbox. Follow the activity and see what changed.
Review before you accept.
Inspect the patch, check the tests, and preview generated websites. The final decision stays with you.
01 / BUILT FOR THE WHOLE JOURNEY
One workspace.
Every next version.
A place for your ideas to become working code, with a clear trail from the first instruction to the final result.
A fresh space to work.
Every cloud task gets an isolated Vercel sandbox. The agent can edit files and run commands while your API credentials stay outside the execution environment.
Close the laptop. Keep the work.
The cloud runs the task. Neon keeps its history, events, and exported results available when you come back.
See the change, not just the answer.
Review the actual patch and test output. Open a generated website preview, download your changes, and choose what to accept.
Bring your repository.
Import a branch from GitHub, edit it with the agent, and export reviewed changes to a draft pull request.
Build with a live preview.
Keep the conversation beside your files. Run the frontend and backend together, then publish a saved revision.
Make room for your crew.
Invite editors and viewers. Keep workspace access, monthly budgets, model usage, and saved revisions visible.
ISOLATED EXECUTIONPERSISTENT HISTORYHUMAN REVIEWAUTOMATIC CLEANUP
02 / TAKE A LOOK INSIDE
From “what if”
to “there it is.”
Explore an example run. Switch between the task, the proposed change, and the checks that make it ready for review.
UseBoat / example runILLUSTRATIVE DEMO · NO AI CALLS
THE BRIEF
A small bug.
A small bug.
A clear destination.
Fix the calculator’s addition function. Keep the acceptance tests unchanged and make sure positive and negative numbers work.
calculator.jsnode --test// Before the agent’s change
export function add(a, b) {
return a - b;
}
// Expected
add(2, 3) → 5
THE PROPOSED CHANGE
A patch you
A patch you
can inspect.
See exactly which lines changed before accepting the result. Protected tests stay in place.
1 file changeddiff --git a/calculator.js b/calculator.js
export function add(a, b) {
- return a - b;
+ return a + b;
}
Acceptance tests: unchanged
THE REVIEW
Checks first.
Checks first.
Your decision next.
Successful verification puts a result in the review queue. Acceptance is an explicit decision, not an automatic merge.
Ready for review✓ Positive numbers
✓ Negative numbers
✓ Zero values
✓ Protected tests unchanged
Sandbox: stopped
Result: awaiting your review
Example output for explaining the workflow.Run your own task ↗
03 / CHOOSE YOUR HEADING
What will you
put in motion?
Start with a focused brief. Small, self-contained tasks are the best fit for the current workspace.
01 / BUILD
A website from a brief. ↗
Landing pages, portfolios, and small interactive experiences with inline HTML, CSS, and JavaScript.
02 / FIXA bug with a clear test. ↗
Give the agent project files and an acceptance command. Inspect the fix alongside the test results.
03 / EXPERIMENTA useful little prototype. ↗
Try a calculator, a text utility, or a simple browser game that works without external dependencies.
04 / PROGRESS YOU CAN INSPECT
A result needs
more than a promise.
The workspace records the steps that led to a result. Here is what has been verified in our own cloud release checks.
A real OpenAI run produced an HTML website, passed its acceptance checks, and exported an interactive preview.
A running cloud task was cancelled and its sandbox stopped. Successful runs also clean up after export.
Patches, history, and acceptance decisions were verified through the deployed API and Neon database.
Release checks performed October 6, 2026. These are functional checks, not speed benchmarks or a guarantee that every generated result will pass.
05 / KNOW WHAT YOU’RE RUNNING
Clear boundaries.
Visible costs.
This is a private workspace using your connected services. There is no subscription checkout here. Your providers bill their own usage.
OpenAI
Model usage is billed to your OpenAI account. Up to 30 requests and 8,192 output tokens per request, per task.
Token usage, model estimates, and budget reservations are tracked per task.Vercel
Sandbox compute is shown as an estimate. Function hosting and any provider charges are separate.
The compute budget does not cap your total provider bill.Neon
Task state, events, and exported artifacts are stored in Postgres. Database usage follows your Neon plan.
Projects, conversations, revisions, and team access persist in Neon.Two ways to work.
| Capability | Local development | Cloud workspace |
|---|---|---|
| Execution | Docker on your machine | Isolated Vercel Sandbox |
| Laptop can be offline | No | Yes |
| AI billing | Your OpenAI account | Your OpenAI account |
| Compute billing | Your own hardware | Provider usage; estimate displayed |
| Task time limit | Up to 3,600 seconds | Up to 20 minutes |
| Live cloud concurrency | One local worker | Up to 3 active tasks |
| Outbound sandbox network | Disabled | npm registry during setup; disabled during agent execution |
A QUICK WORKLOAD CHECK
How much task time?
Explore a sample workload. This estimates runtime, not price or completion speed.
total sandbox minutesAssumes every task runs for the full selected duration.
06 / BEFORE YOU SET SAIL
Good questions.
Straight answers.
Do I need to sign in to explore?
No. This landing page and its demo are public. Sign in is only required for the workspace, where tasks use your connected API account and saved project data.
Does it keep working when my laptop is off?
Yes, cloud tasks run on Vercel and save their results in Neon. Tasks run for up to 20 minutes using saved checkpoints. If a step is interrupted, you can review its status and explicitly continue from the last saved files.
Can it build websites and games?
Build websites and small browser experiences with multiple files, npm dependencies, and Node backends. Preview static, Vite, Next.js, or Node projects in a separate sandbox. Previews last 15 minutes; publishing requires a connected Vercel account.
Can I bring my own code?
Yes. Import a GitHub repository and branch, or edit files directly in Studio. Projects support up to 200 text files and 2 MB; GitHub imports allow 198 files and 1.8 MB. npm installation is restricted to the package registry, with lifecycle scripts disabled.
What happens after a task finishes?
Review the patch and acceptance tests, then accept the result as a new saved project revision. Continue with another instruction, preview the app, create a draft GitHub PR, or publish the saved revision to Vercel.
Is the compute budget my total spending limit?
Set separate model and compute limits for each run, plus a monthly workspace budget. Studio reserves estimated costs before work starts and displays usage alerts. Provider invoices, hosting, database, and published-app charges remain separate.
Can my whole team create accounts?
The owner can invite teammates as editors or viewers using one-use links. Each account sees only its workspaces. Members can change passwords, and the installation owner can issue account recovery links.
Is this the official boat.dev service?
No. UseBoat is an independently built workspace inspired by nautical exploration. It uses OpenAI, Vercel Sandbox, and Neon, and is not affiliated with boat.dev.
Let’s see where
this goes.
Give your agent a destination.
Come back with something you can
review.