Security model

End-to-end encryption

Messages are encrypted before they leave the browser or local agent. As an agent builder, avoid logging decrypted user content unless the user explicitly expects that behavior.

Device trust

Each trusted browser has local state. Clearing browser data, switching profiles, or using a private window can make Pacerelle ask the user to sign in again.

No inbound ports

Local agents connect outward from the user's machine. You do not need to expose a public local HTTP server just to receive Pacerelle messages.

Agent token security

Agent tokens authenticate local runtimes. The token is shown once when an agent is created or token-rotated.

Do:

• store the token in .env, Docker secrets, or a secret manager
• rotate the token when it is exposed
• restart the local runtime after rotation

Do not:

• commit the token to Git
• paste the token in client-side browser code
• log the token in CI output
• share the token across unrelated machines

Local logs

Local logs are part of your security boundary. Log connection state, errors, and timing, but avoid logging message text, file contents, tokens, or widget answers in production.

Attachments

Treat attachments as sensitive user input. Validate type and size before processing, and ask for confirmation before sending file content to another service.

Production checklist

• Rotate leaked agent tokens immediately.
• Keep decrypted logs disabled in production.
• Use confirmation widgets before file writes, command execution, or network calls.
• Validate attachments before processing them.
• Prefer short-lived credentials for downstream tools.
• Make token rotation part of incident response.
• Avoid running local agents with broad filesystem permissions unless the workflow needs them.
• Separate development and production agents.