Your agent stays on your laptop
Add the plugin and OpenCode gets sandbox tools. The conversation and your model keys stay on your machine.
Cloud sandboxes for the agent on your laptop, or a cloud workspace for the whole agent. From the provider you choose, and you can switch any time.
OpenCode on your laptop runs tests, builds and previews in cloud sandboxes.
Then fill in your config. No Gitterm account needed.
OpenCode, Codex or Claude Code in a cloud workspace. Reach it from your phone or Slack.
Then sign in and type /cloud, or start one in the browser.
Extend your laptop
Your agent can attach sandboxes from any of these. The demo moves through them on its own, or pick one to see it.
Sandboxes attach as the agent needs them.
Your Mac ran out of memory mid-run.
This code expired on 31 October.
Restored here. Tests pass and the checkout preview is up.
Opened PR #482 from checkout-coupons.
To avoid problems with your computer, quit any applications you are not using.
Tests, Postgres and the dev server are all running on this Mac.
Your teammate can’t reach “localhost:3000”. It only exists on your Mac.
OpenCode, Postgres and the dev server stop when your Mac sleeps. The session can’t come with you.
Add the plugin and OpenCode gets sandbox tools. The conversation and your model keys stay on your machine.
The agent starts a sandbox on the provider you set and syncs your working tree. Ignored files like .env stay home. Once it has saved a setup snapshot, the next sandbox starts with dependencies already installed.
sandbox_createsandbox_sync
Tests, builds and dev servers run in the sandbox. The agent gets the output when a job finishes, or checks back on a long one, while your laptop stays idle.
sandbox_execsandbox_job
A port becomes a preview URL you can open. Files changed in the sandbox only land in your checkout when the agent applies them.
sandbox_previewsandbox_apply
Add the cloud plugin and type /cloud. The conversation, your unpublished commits and, if you choose, uncommitted changes move to a cloud workspace. Pick it up from your phone, a browser or OpenCode desktop. Your laptop copy stays as it was until you bring the session back with /cloud-return.
limits to any of them, so a busy agent cannot run up a surprise bill.No Gitterm account needed. The key comes from your environment and the provider bills you directly.
"options": { "provider": "e2b", "apiKey": "{env:E2B_API_KEY}" }
No provider account needed. Sign in with gitterm login; sandboxes count against your plan, and Gitterm stops anything that outlives its deadline.
"options": { "size": "medium" }
Point at a self-hosted server. It uses its default provider; add provider to pick another one it offers.
"options": { "serverUrl": "https://git.acme.dev" }
/cloud pairs to the app with a one-time link.Run agents in the cloud
/cloud, or start Codex or Claude Code in a new one from T3 Code. They keep running on their own, and your laptop is free for whatever is next. /cloud-return brings a session back whenever you want it.OpenCode
Move a session with /cloud, back with /cloud-return
T3 Code
RunsCodex and
Claude Code
Move an OpenCode session to the cloud, keep the agent going on your phone, in a browser or on another computer, then pull the session back to your laptop whenever you like, finished or not.
The whole agent session travels: its conversation, and the checkout it works in, with commits and uncommitted changes. Anything you changed on your laptop in the meantime is kept.
Keep a snapshot as you return, and the next /cloud starts with the setup done.
Scan a QR code to open the workspace in the OpenCode app on your phone, or open it in any browser. Pause it when you are done and pick up later.
ws-7k2.gitterm.dev/auth/connect/7QK4M25 minIdle workspaces pause on their own.
Mention @gitterm in a channel. It works in a cloud workspace for that repo and answers in the thread. Self-host the bot from one Docker image.
@gitterm the checkout test is flaky again, can you take a look?
On it. Working in acme/shop in a cloud workspace.
Fixed the clock dependency. 20 runs passed. @Ada
The SDK creates a workspace, starts a run and waits for the result. Use it from scripts, CI or your own tools.
const { workspace } = await client.workspaces.create({ repo: "https://github.com/acme/shop" }); const run = await client.runs.create({ workspace, prompt: "Fix the flaky checkout test." }); const { finalText } = await client.runs.result(run);
/cloud make one for you.Stop paying for compute, keep the workspace, pick up where you left off.
Expose any port as a private URL, or make it public for webhooks and demos.
Open it in your local OpenCode TUI with one command, or in the T3 Code desktop app.
Same workspace in your browser, OpenCode desktop or your phone.
Model keys are saved once and reused in every workspace, never pasted again.
Create workspaces with the SDK. Manage them with the CLI or the SDK.
Add the plugin to OpenCode, or start a cloud workspace. Open source and self-hostable.