Run Claude Code in a sandbox

Claude Code (claude) is installed in the ubuntu template. Choose how it signs in, then run a task with claude -p.

Sign inCopy Anchor Link

With an Anthropic API keyCopy Anchor Link

Put the key in the env of the sandbox:

{
  "ANTHROPIC_API_KEY": "sk-ant-..."
}

Claude Code bills the usage to that key. See authentication in the Claude Code docs.

With a Claude subscriptionCopy Anchor Link

Sign in inside the sandbox, so that the sign-in never passes through the API:

1. Connect over SSH and run claude.
2. Type /login and choose your Claude account. Claude Code prints a link.
3. Open the link on any device, sign in with your Pro or Max plan, and paste the code back into the terminal.

Claude Code keeps the sign-in in the sandbox, and each later claude -p in the same sandbox uses it. A persistent sandbox keeps it across pauses. Claude Code then uses the usage limits of your plan.

Anthropic sets the limits of a plan for the usual work of one person, and it does not allow a third party to collect or store the tokens of a Claude account (legal and compliance). So do not put such a token in an API request, and use an API key to run many agents at the same time.

With the open models of Munito AICopy Anchor Link

{
  "ANTHROPIC_BASE_URL": "https://api.munito.ai/inference",
  "ANTHROPIC_AUTH_TOKEN": "munitoai_...",
  "ANTHROPIC_MODEL": "zai-org/glm-5.3-flash[1m]",
  "ANTHROPIC_DEFAULT_OPUS_MODEL": "zai-org/glm-5.3-flash[1m]",
  "ANTHROPIC_DEFAULT_SONNET_MODEL": "zai-org/glm-5.3-flash[1m]",
  "ANTHROPIC_DEFAULT_HAIKU_MODEL": "qwen/qwen3.8-27b"
}

One Munito AI key with the Inference permission serves the sandboxes API and the models. Claude Code with open models explains each variable, the models and the regional hosts.

Run a taskCopy Anchor Link

curl -sS --fail-with-body "https://api.munito.ai/sandboxes/v1/sandboxes/$SANDBOX/exec" \
  -H "Authorization: Bearer $MUNITO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"command": "cd /workspace && claude -p \"Write a Python script that prints the first 10 primes, and run it\" --allowedTools Bash Edit Write Read --output-format json", "timeout": 1800}'

• claude -p runs one task and exits. --output-format json returns the result and the session_id.
• --allowedTools Bash Edit Write Read lets Claude Code run commands and change files without asking. As root, Claude Code refuses --dangerously-skip-permissions, and --allowedTools does the same work.
• To follow the work while it runs, use --output-format stream-json --verbose with "stream": true in the command: each event of Claude Code comes as it happens.
• claude -p "..." --resume <session_id> continues a session in a later command.
• A command runs for at most 30 minutes. For a longer task, run nohup claude -p "..." --allowedTools Bash Edit Write Read > /workspace/claude.log 2>&1 & and read the log with later commands.

Use the terminal UI over SSHCopy Anchor Link

Add your public key with SSH, connect with the command that GET /sandboxes/{sandbox_id}/ssh answers, and run claude. A shell starts in /workspace, with the variables of the sandbox. For a session that outlives your connection, install tmux (apt-get install -y tmux) and start Claude Code inside it.