# Frequently asked questions

Short answers to what people ask before their first run.

### Does my code leave the sandbox?

Only the diff, and only into a pull request on your repo. Outbound network is off unless you [allow a host](/sandbox/network), and the sandbox is wiped when the run ends.

### Do you train models on my code?

No. Runs on Team and Enterprise are covered by a zero-retention agreement with every model provider we route to.

### Which agents can I use?

Claude Code, Codex, or any process that speaks MCP over stdio. See [Connect your agent](/agents/connect).

### What happens when a run can't pass the evals?

It stops after three tries with the status `needs_review` and no pull request. The trace shows the failure the agent couldn't fix. See [Eval gate](/review/eval-gate).

### Can Kestrel merge its own pull requests?

No. Kestrel opens pull requests and answers review comments. Merging follows your branch rules, like any other pull request.

### How much does a run cost?

A typical run on Team takes about 11 minutes and costs $1 to $3 in usage, most of it sandbox time. Each run's cost is on its page in the dashboard and in `run.wait()`. See [Billing & limits](/billing).

### Can I run Kestrel in my own cloud?

Yes, on Enterprise. Sandboxes run in your VPC, with your own model keys, and nothing leaves your network.