BENCHYARD GUIDES
How Benchyard workspaces work
A workspace is where your agents work: it holds your project's files, your sessions and the tools they use. You reach it from the browser on your computer or phone.
Last reviewed . What is available depends on your installation and account.
What persists when I leave?
Closing a browser tab or your laptop does not stop a running workspace. An agent keeps working until it finishes, needs you, fails, or the workspace stops.
Stopping a workspace ends its terminals and running programs. Its files, conversations and sign-ins are kept. A machine can also be suspended: what was running is kept and carries on when it starts again. A container is stopped, not suspended. The workspace's card says which happened and why.
Opening a session of a stopped or suspended workspace does not start it. Sending a message does: the message is held and sent once the workspace is up, and you can take it back before then. Opening the terminal, the files, the desktop or a preview also starts it.
A workspace is stopped when idle only if you ask for it. The setting is Never unless you choose 30 minutes, 1 hour, 4 hours or 1 day; a size of workspace can have a time of its own, and Keep running exempts one workspace. Idle means no agent is at work or waiting for you, nothing is running in its terminals or in the background, and you have not had it open for a while.
Where does the workspace run?
A machine type is the size of a workspace: how many processors and how much memory it has. Shared types share processing power with other workspaces, so their size reads “up to”; Dedicated types have their processors and memory reserved. A machine type also says whether it has a desktop. Which types you can choose depends on your installation and account.
| Kind | What it is |
|---|---|
| Container | A light, isolated environment for each workspace. When stopped, what was running ends and its files are kept. |
| Machine | A virtual computer for each workspace, with its own disk. It can be suspended when idle, so that what was running carries on when it starts again. |
What can I do inside a workspace?
The terminal in the browser keeps going when you close the tab, and you can open the same terminal from several tabs or devices. The file browser shows the project's folder, edits text files, and uploads or downloads one file at a time.
To use your own terminal or editor, connect over SSH: install the benchyard command on your computer, sign in with it, add your public key in the app, and connect to the workspace by its name. The usual tools for copying files work too. The connection is encrypted and passes through Benchyard's server. Like the rest of Benchyard, it is not end-to-end encrypted.
Where a workspace includes a desktop, it opens in the browser. It has no sound, and programs that need a graphics card run slowly. Only the workspace's owner can use any of these.
How do snapshots and rollback work?
A snapshot is a saved copy of the workspace's files at one moment. It does not hold what was running. You can take up to ten yourself, and choose a daily one that is taken while the workspace is idle or stopped. Only the workspace's owner can use them.
Going back to a snapshot (a rollback) asks you to confirm, saves the present state as a snapshot first, and then puts the files back as they were. Sessions and messages from after the snapshot leave the history. Rolling back to the saved state undoes it. You can also make a new workspace from a snapshot.
Where a snapshot is kept depends on the installation's storage, and the app says which. On a cloud provider's storage it is kept separately from the workspace's disk, so it survives if the disk is lost. An installation whose storage keeps snapshots alongside the disk says so: there a snapshot undoes mistakes but is lost if that storage is. Whether snapshots are offered depends on the installation.
How do I continue on my phone?
Sign in on the other device and open the same workspace and session. One device drives a session at a time, and the others can watch and approve. To carry on from another device, just send a message there: that device takes the session over, and a line in the session says which device is driving.
On an iPhone, add Benchyard to the Home Screen before turning on notifications. You turn them on for each device, and the device asks your permission. A notification opens the session, where you look at the request and approve it.
How do browser previews work?
Where previews are available, the Ports list offers Open in browser for an app running inside a workspace. The preview opens at an address of its own, separate from the Benchyard app.
A preview opens only for the workspace's owner, signed in. It is not a public link or a way to share the app with other team members. Opening a preview starts a stopped workspace, but the app itself has to be started again before there is anything to see.
What is pushed, and what is emailed?
No email is sent while you have the session open, and at most one per session in ten minutes. Delivery of email and notifications is not guaranteed.
| Channel | What to expect | Requirements |
|---|---|---|
| Device push | An approval is needed, a plan is ready, a task finished or failed, a note from the agent, a workspace is ready. | Turn it on for each device. On an iPhone, add the app to the Home Screen first. |
| An approval is needed, a plan is ready to review, a task failed, an agent needs you to sign in again. Not for finished tasks or agent notes. | Turn it on and confirm your email address. The installation must have mail set up. |
The demo, step by step
The demo on the homepage is a replay drawn for the website, not the live app. The changes, files, test results and issue board in it come from a real run in Benchyard with demo data; two lines of unrelated notices are left out of the test output. The agent's replies are scripted.
Play the demo on the homepage.
- 1
A task is sent.
In the issue-board workspace, Alice asks Codex to add a status filter to a fictional issue board and test it. The session is in Ask first mode.
- 2
The agent edits three files.
Codex changes three files. Each edit appears in the conversation and under Changes, where you can see exactly what changed.
- 3
It asks before running the tests.
The session shows Approval needed for the command npm test and waits. The same request appears on a signed-in phone, which can watch and approve while the MacBook drives.
- 4
You decide.
Approve, and the command runs: four filter tests pass. Deny, and it is not run; the replay lets you ask for the request again.
- 5
You look through the result.
Changes shows what changed in each of the three files, Files the project's files, and Terminal the recorded output of the command. Ports lists the app running in the workspace.
- 6
You open the running app.
Open in browser shows the issue board. In progress and Done each show three issues, All issues shows six.