Workspaces
A workspace is a sandbox with a persistent folder. You start it from an agent, talk to the agent in a chat and stop it again. Files stay when it stops; an idle workspace stops by itself after 30 minutes. To change that, or to keep a workspace running until you stop it, use Settings → Stop when idle (the workspace’s creator or an admin).
Starting a workspace
Section titled “Starting a workspace”Choose New workspace, pick an agent and, if you want, a name. Without a name the workspace is named after its agent. The agent holds the whole configuration: engine, instructions, skills and access. A workspace only runs it, so start as many workspaces from one agent as you need, and change the agent to change all of them at their next start.
The sandbox is ready after a few seconds. It runs on your credentials and your access. To pin a specific engine credential, or to give one workspace more access than its agent, use Integrations → Edit access (agent managers).
Replies stream in as the agent writes them. Below the message box you choose the permission mode and the model of the next message. The lists show only what the workspace’s engine supports:
| Mode | Claude Code | Codex | OpenCode |
|---|---|---|---|
| Ask before acting | yes | yes | yes |
| Plan only | yes | no | yes |
| Accept edits | yes | yes | yes |
| Auto (classifier) | yes | no | no |
| No questions asked | yes | yes | yes |
You can write while the agent works. While it is busy, your message goes into the running turn and shows Queued until the agent has read it. While the agent waits for your answer to a question or permission, the message becomes the next turn: it is listed above the message box and starts by itself when the current turn ends. Cancel removes it before it starts.
Until you choose, the agent’s defaults apply. Your choice is remembered for the workspace in this browser. OpenCode lists the models its credential serves while the workspace runs.
To give the agent files, use the paperclip, paste them or drop them on the
message box (at most 25 MB each). They are stored in /workspace/uploads,
and the message tells the agent their paths. A file never replaces another:
when the name is taken, a short hash of the content is added
(image-3f2a9c1b.png).
A conversation keeps its context when you stop and start the workspace: the
engines keep their sessions in the workspace folder (.delegate/state, hidden
in the file browser and part of a snapshot).
Integrations
Section titled “Integrations”The Integrations button shows what the agent can reach and whether it is ready for you. An integration that needs your own account shows Not connected; connect it under My connections and start the workspace again.
Activity
Section titled “Activity”The Activity button lists every outbound call of the sandbox: what was called, whether it was allowed, and which rule decided.
What the agent may do
Section titled “What the agent may do”Reads are allowed. Calls that change something are refused with “change set required”. The agent then proposes a change set for a person to approve.
Snapshots
Section titled “Snapshots”The Snapshots button copies the files of a stopped workspace to the organization’s object storage, and restores them into a stopped workspace. A restore replaces the files that are there.