Agent kits

An agent kit is a sandbox kit with an AI coding agent already installed and pinned to a tested version. You launch it, open the web shell, and the agent is ready, with the API keys you chose already in its environment. Each kit sits on the same Debian 13 base with git, Python and Node, so the agent can clone a repository and run code in an isolated machine instead of on your laptop.

Before you begin

• Your provider has installed the agent kits. If they have not, the launcher says “No agent kits are installed” and starts with an operating system kit; ask your provider to install them.
• An API key from the agent’s vendor (see the table), saved as a secret or ready to paste.

The kits

The Kits page lists them under the Agents filter.

Kit	What it is	Key it asks for
Claude Code	Anthropic’s coding agent, ready in the web shell.	Anthropic API key
Codex CLI	OpenAI’s coding agent in your terminal.	OpenAI API key
Gemini CLI	Google’s Gemini agent in your terminal.	Gemini API key
OpenCode	The open source coding agent (1.x line), bring your own model.	Fireworks, OpenRouter, Anthropic, OpenAI, Gemini, DeepSeek or Kimi API key
OpenCode 2	The OpenCode 2 rewrite of the open source coding agent, bring your own model.	The same provider keys as OpenCode
GitHub Copilot CLI	GitHub Copilot’s terminal agent.	A GitHub token with the Copilot Requests permission
Shell	A plain Linux shell with git, Python and Node, and no agent.	None

Every kit also accepts a GitHub token for repository access (see The GitHub token). Agent CLIs are installed from each vendor’s own channel when your provider builds the kit, at a pinned version, so a vendor update never changes a kit underneath you. New kit versions are installed by your provider; the launcher’s kit picker lets you choose an older version if you need one.

Launch an agent

1. Go to Sandboxes > Sandboxes and click Launch.
2. Click the kit in the sentence and pick the agent kit.
3. Click the secrets part of the sentence (the with part) and tick the API key for that agent. See Secrets and API keys.
4. Pick a location and size, and a lifetime. Agent kits default to 1 hour.
5. Click Launch, then open the sandbox’s Shell tab and start the agent from the prompt.

Secrets and API keys

A secret is a named value, such as ANTHROPIC_API_KEY, kept encrypted in your account. When you launch, the secrets you tick become environment variables inside the sandbox, so the agent finds its key without you pasting it each time.

Where secrets live

Go to Sandboxes > Secrets to see, add, replace and delete your secrets. A secret has a name (capital letters, digits and underscores, like OPENAI_API_KEY) and a value. The value is write-only: the panel never shows a saved value back to you. You can save a secret from the launcher too, without leaving the page.

Choosing secrets in the launcher

Click the with part of the sentence. The popover lists:

• Provider rows for the keys this kit suggests, each with the vendor’s mark. Tick a saved key to use it, or add one in place. A provider row is optional: “Launches without it; you can also set it inside the sandbox later.”
• Other saved secrets, anything else in your account that you want in this sandbox.
• A replace link on each saved secret. Replacing changes the secret for every sandbox you launch afterwards on your account (a sandbox that is already running keeps what it was started with), and the panel warns you before it does.

Custom keys and values

For a value that is not a saved secret, open Advanced options > Environment and add key-value pairs. They are merged with the kit’s own defaults. A key you set there overrides the kit’s value, and it also overrides a picked secret with the same name.

When they take effect

Secrets and variables are written into the sandbox when it is created. Changing them later does not reach a sandbox that is already running; start a new run for the change to take effect. They are visible in a login shell (the web shell, bash -l), but not to a non-login command such as ssh host cmd. If the node running your sandbox is too old to deliver them, the launch is refused with a message saying so rather than starting a sandbox without its keys.

Caution

Anything running inside a sandbox can read its environment, including the agent and any code it runs. Give a sandbox the keys it needs and no more, and prefer keys you can revoke.

The GitHub token

The last line of the launch sentence reads Add [a GitHub token] for repo access., or Using [your GitHub token] for repo access. once one is picked. It shows for every kit, because git works the same way in all of them.

1. Click the token in that line.
2. Paste a personal access token, or choose Use my saved GitHub token. Fine-grained tokens work; grant them access to the repositories this sandbox needs and no others.
3. The token is stored encrypted as the secret GITHUB_TOKEN and reused for every kit.

Inside the sandbox, git signs in to github.com with that token automatically, so git clone and git push over HTTPS work without a prompt. The helper is registered for github.com and sends the token to that host. Replacing the token changes it for every sandbox you launch afterwards.

Registry credentials for private kits

If you build a kit from a container image in a private registry, the build needs a pull credential. It is saved as a registry secret and picked on the kit’s Private registry credential field; it is never given to a sandbox as an environment variable and does not appear in the launcher’s secrets list. The Secrets page lists registry credentials separately. For a Docker Hub credential it offers Check pull limit, which shows how many pulls remain in the current window as seen from your provider’s panel. See Private registries.

Tools for the agent

Agent kits can be given external tool servers through the tools gateway. See Tools (MCP).

Common problems

• The agent asks you to sign in even though you picked a key. The key was not ticked at launch, or it was added after the sandbox started. Secrets reach a sandbox when it is created; start a new run.
• The agent says the key is invalid. Check the key at the vendor, then use replace on the saved secret and launch again.
• git clone of a private repository asks for a password. No GitHub token was picked, or it does not have access to that repository.
• GitHub Copilot CLI cannot authenticate. It needs a fine-grained personal access token with the Copilot Requests permission, given as the Copilot token; an ordinary classic token does not work.
• The launcher opened on an operating system, not an agent. No agent kit is ready yet: they may still be building, or your provider has not installed them.

Related

• Launch a sandbox
• Tools (MCP)
• Create a kit
• Web shell, logs and metrics