Security

Isolated by design

An agent that can run commands needs walls around it. Here is how Windlass keeps accounts apart and your data yours.

One container per account

Every workspace runs in its own container with its own filesystem, process space and network namespace. The agent's shell cannot see other customers' containers or the host.

No model keys in the workspace

Workspaces reach the model provider only through the Windlass gateway, using a per-account credential that carries no access to anything but your own allowance. A leaked credential can be rotated in seconds.

Encrypted in transit

All traffic between your browser, Windlass and the model provider uses TLS. Sessions are cookie-based, HttpOnly and same-site.

Access controls

Only you can open your workspace and files. Support staff do not access workspace content unless you ask for help with a specific issue.

Retention you control

Workspace content stays while your account exists. Delete the account from Settings and content is removed within 30 days, backups within 30 days after that. Export files any time from the Files page.

Metering, not surveillance

Usage records contain token counts, the model and timestamps — not prompt text. Technical logs are kept for 90 days.

Where your data goes

  • Your workspace — files, session history and attachments live on Windlass servers (our hosting provider).
  • Model inference — prompts, the files you send to the model and the model's replies are processed by DeepSeek's API (Hangzhou DeepSeek AI, China) under DeepSeek's API terms. We do not use your content to train models.
  • Voice — recordings are transcribed by a speech-to-text provider and not retained after transcription.
  • Billing — payment details are handled by the payment processor; we never see card numbers.

Full list of sub-processors in the Data Processing Addendum.

Good practice for agent work

  • Give connectors scoped tokens with the least access needed.
  • Do not paste production secrets into the chat; put them in environment variables or files the agent reads.
  • Review commands and diffs before you rely on them; the agent is fast, not infallible.
  • Keep one project per client or codebase so context never mixes.

Found a vulnerability? Write to our security contact. We acknowledge reports within two business days.