Data location
On the cloud service, your data is kept in our control plane, your code passes through the execution host that runs each turn, and what the agent reads goes to the models we run. In a private deployment everything stays inside your network.
Cloud service
| Data | Where it is |
|---|---|
| Sessions, messages, settings, ledger, audit log | The control plane's database |
| Artifacts: patches, changed files, test output, reports, memory drafts | The control plane's object storage |
| Your repository's code | On GitHub. For each turn, the control plane fetches the pinned commit and passes it to the execution host. That copy is gone when the sandbox is destroyed. The changes are kept as artifacts |
| What the model reads: your message, code the agent looks at, command output, retrieved passages, project memory | Sent to the model deployment that the turn uses, which we run |
| Knowledge documents | Your organization's account in the knowledge service we run. Their text is sent to an embedding model we run, for indexing |
| Calls to your MCP tools and retrieval services | Sent from the control plane to the endpoints you configured |
| Published changes | Pushed to your repository on GitHub |
What a project's region limit does
A project can be limited to certain clouds and regions. The limit decides which execution pools may run its turns, and so where the code is checked out and the sandbox runs. It doesn't move the control plane or the stored data.
What happens during a turn
- The control plane resolves the branch to a commit and pins the turn's configuration: role, model, memory, knowledge sources, tool connections and rates.
- A host in the chosen pool starts a sandbox and gets the source at that commit through the control plane, with a short-lived credential for this turn only.
- The agent works in the sandbox. The sandbox has no network access. Model calls and tool calls go from the trusted side of the host through the control plane's gateway.
- When the turn ends, the patch, test output and report are uploaded to the control plane, and the sandbox and, for a fresh virtual machine, the machine are destroyed.
Private deployment
Everything stays inside your network: the control plane, its database and artifacts, the knowledge service, the execution hosts, and the identity provider, Git server, models and MCP tools you connect. Online payment, fallback to public models and SDK telemetry are switched off. See Private and offline deployment.