Security

• Every record is tied to the agency it belongs to, and the database filters every console read and write with row-level security (RLS) policies: a person only sees rows of the agencies they are a member of.
• The agent identifies the agency by the number the client wrote to, verified with Twilio's signature, and every query it makes is scoped to that agency.
• On every code change, automated tests verify that one agency cannot see another's data.
• Tango's team sees a directory of agencies (name, owner email, line and onboarding status) to onboard them, not their conversations, clients or listings. If the agency asks for it when signing up, a Tango team member joins as a support admin: they appear in the team and the agency can remove them at any time.

• Four roles: owner, admin, agent and viewer. Owners and admins manage the team, listings and settings; agents handle conversations; viewers can only look.
• Every permission is enforced in three layers applying the same rule: the database, the server and the interface.
• People join a team only by invitation —from an owner or admin, or the onboarding invitation Tango sends the owner—, and their email is verified before access is granted.
• Sign-in is by email and password or with Google. Passwords are stored hashed by the authentication provider: we do not store them.
• The console always acts with the person's own session: it holds no credential capable of bypassing the database's permissions.

• All traffic is encrypted with TLS: between your browser and the console, between our services and with every provider.
• The database and photos are stored encrypted at rest by our infrastructure providers.
• Each agency's WhatsApp (Twilio) access token is encrypted with a public key as soon as it is entered, and only the agent can open it. The encryption is bound to the agency: copied to another one, it is useless.
• Production secrets live only in each provider's production environment, never on personal computers.

• Every incoming message is verified against Twilio's signature before it is processed, with a per-sender rate limit; a repeated message is never processed twice.
• The agent can only read the listings, never modify them, and is instructed to talk only about properties a search returned in that conversation.
• When a team member takes over a conversation, the agent stops replying: the database guarantees the two never answer at once.
• On WhatsApp, a viewing offer can only be taken by the member it was sent to, from their verified WhatsApp number and on their own agency's line; in the console, by a member with their session and role. The number is verified by sending from it a code the console shows only to that member, and an offer forwarded to and tapped from another phone is not taken by anyone.
• If something fails during a turn, the client gets a safe reply instead of silence, and the agency's handoff contacts are alerted.
• Replies are generated by an Anthropic model through its commercial API, which does not use that data for training.

• Technical logs are designed not to keep message content or client data: phone numbers are masked and errors are recorded by type and code.
• In the agent, database errors are logged by code, without the detail that could carry a row of data.

• The database is hosted in São Paulo, Brazil. The full list of providers is in the Privacy policy.
• Development and production are separate environments, with different databases and credentials. The exception is property photos, delivered through a public URL that cannot be guessed: they live in a single Cloudinary account, in folders separated by environment, and each environment's key reaches the whole account. Tests run against local databases, never against real data.
• No change reaches production directly: it goes through a pull request, automated tests and the development environment, and a person releases it.
• Dependencies pinned by version and hash, secret scanning on every change and least privilege at every deployment step.
• Database permissions are checked automatically: no table is left exposed without row-level security.

If we detect an incident affecting personal data, we contain it, investigate it and notify the affected agencies without undue delay with what we know: what happened, what data it reached and what we are doing about it. Where required, we also report it to the supervisory authority.

If you find a security flaw, write to us at santino.fajardo@tangoagent.app with the details and how to reproduce it. We will reply, let you know when it is fixed, and we will not take action against good-faith research that does not access third-party data or disrupt the service.

• Give each person the least role they need, and use “viewer” for anyone who only needs to look.
• Remove people who stop working with you from the team: they lose access immediately.
• Use unique passwords, or sign in with a Google account that has two-step verification.
• Tell your clients they are being helped by an automated assistant.