Launch your first session

To launch a coding agent in an isolated sandbox you describe how you work once, as a workspace, and then launch a session of it. The session is a fresh container on one of your organization's servers with the repo cloned on its own branch, the environment loaded and your agents inside.

Create a workspace

A workspace is yours. Teammates cannot see it, and each session of it works on a branch that merges back into your tree. Open Workspaces, choose New workspace, and fill in three things.

One repo, one terminal and one agent view is a fine first workspace. The full editor is described in repos, views and environments.

Pick an AI credential

Isla, the starter agent, arrives with your membership: a careful senior engineer who reads before writing, works in small verifiable steps and never prints a secret. She arrives without a credential, and until you assign her one she cannot answer anything. On the Credentials page add an AI credential, either an API key or a subscription such as Claude Pro or Max. The harness follows the credential: Claude Code for Anthropic, Codex for ChatGPT, goose for any other API key. Credentials are personal, and the real key never enters the sandbox; a gateway injects it on the way out. The options are in AI providers.

Launch

Press Launch on the workspace. If it has several environments you pick one; if the organization has several servers and no default, you pick a server, each labeled with whether it is up. A sleeping Cloud Server wakes first, which can take a few minutes.

Then, in order: the server creates a sandbox from the workspace's devcontainer image; the repos are cloned, credentials never in argv or a URL, and the session starts on its own session/<id> branch; the environment's variables and secret files are unsealed and materialized inside the sandbox, files at mode 0600, for the sandbox's lifetime only; each view's process starts, the terminal as ttyd over tmux, the agent view as a real harness session. The session goes from creating to ready and the views appear.

Open the terminal and talk to Isla

The terminal drops you in /workspace with the repo already there and your variables in the environment. The agent view is a conversation with one agent: streaming replies, tool calls, diffs you can read before they land, permission prompts for anything destructive, and slash commands. Try something concrete:

Read the README and tell me how to run the tests. Then run them.

Isla reports what actually happened, failures included. She keeps a memory per workspace and agent, so what she learns about this repo is there next session. Every view type is covered in views: terminal, code, files, web, agent.

Save, then finish

Nothing lives in the sandbox. When the change is right, press Save: the session's changed files are committed on the session branch and merged back into the workspace's tree, which is stored as an encrypted git bundle with compare-and-swap. You can save a selection of files or everything. Finish then destroys the sandbox. If there is unsaved work the screen says so first, because after Finish it is gone. You can also suspend a session and resume it later, as described in suspend, resume and finish. Every session measures its cost, tokens and time, so the first one already tells you what it cost.

The same launch from Slack

Once the organization installs Slack and you connect to it, the whole sequence is one command in a channel:

/launch <workspace> in <environment> on <server> as <name>

Only the workspace is required; a missing environment or server comes back as a row of buttons. Launching in a conversation connects that conversation to the session, so Isla answers there. /views lists the windows, /view web 3000 publishes a preview, /save and /stop do what the buttons do. How to install and connect it is in Slack.