Trust and security

Security practices

• Encryption

  • All traffic to the console, the API and the voice server uses TLS; the console sends HSTS, a content security policy and frame protection on every response.
  • Provider API keys, webhook signing secrets, customer id secrets and tool header values are sealed with AES-256-GCM before they are stored. Keys are derived per purpose and bound to the organization they belong to, and they can be rotated without downtime.
  • Calls to model, speech and telephony providers go over HTTPS or secure WebSockets.

• Access control

  • Every organization has three roles (Admin, Billing manager, Member) from one permission matrix that every API route enforces; the console disables what a role cannot do.
  • Sign-in supports two-step verification; the security overview shows admins without it, and old or unrestricted API keys.
  • API keys are scoped read or write, can be limited to IP addresses, and test keys cannot delete data or change compliance settings.

• Tenant isolation

  • Every record belongs to one organization, and each request is checked against the caller's membership of the organization it names.
  • Cross-site writes to the API are refused, and outbound requests (tools, webhooks, server URLs) cannot reach private networks.
  • Call tokens are single use and bound to the organization that owns the assistant.

• Activity log

  • Every change is recorded in an append-only activity log with the person, their session, the fields that changed and the outcome; secrets are redacted.
  • Admins can search and export the log, and choose how long it is kept.

• Retention and erasure

  • Each organization sets how long transcripts, recordings and logs are kept; a daily sweep deletes what is older. Recordings default to 90 days.
  • Admins can erase a caller's data on request: calls, recordings, contacts, log mentions and recordings held by the phone carrier.
  • Deleting an organization deletes its data and its stored files.

• Personal data redaction

  • Assistants can redact personal data (such as card numbers, emails and phone numbers) from transcripts, tool arguments and results, workflow values, logs and webhooks.
  • Caller text, knowledge snippets and tool answers reach models only as labelled data, which limits prompt injection.

• HIPAA

  • Healthcare customers sign a business associate agreement in the console; HIPAA mode on an assistant needs an active agreement that covers voice calls.
  • HIPAA mode keeps transcripts, recordings and caller details out of logs, webhooks, exports and analysis, and turns off features that would share them.
  • Video agents run on HIPAA assistants only when the agreement covers video.

• Data residency

  • Organizations choose the United States or the European Union. EU organizations keep recordings, knowledge files and brand voice audio in EU storage and use an EU speech endpoint.
  • Settings and call records are stored in the platform database of the deployment; ask us about a dedicated EU deployment.

• Vulnerability handling

  • We welcome reports from security researchers: write to the security contact below. We acknowledge reports within three business days and do not pursue good-faith research.
  • Dependencies are kept current, inputs are size- and shape-checked, and webhooks are signed so receivers can verify them.

Data residency: organizations are hosted in the United States; the European Union region is available when you create an organization.

Subprocessors

Third parties that may process customer data to run the platform. We give notice on this page before adding one.

Services you connect with your own account are your choice and not our subprocessors: HubSpot, Salesforce, Pipedrive, Google Calendar, Cal.com, Slack, Microsoft Teams, Zapier, Make and n8n, Square, Jobber, NexHealth.

Status and uptime

A machine-readable health check is at /api/health.