Security

Access control

Who can see and change what, how access is enforced, and what the audit log records.

GrantReaches
Workspace membershipEvery agent in the workspace.
Membership limited to chosen agentsThose agents only, and nothing workspace-wide.

Workspace membership

Members of a workspace can read its agents, and members with an editing role can change them. Membership carries a role that governs what the application permits (inviting members, managing billing, changing workspace settings, creating webhooks).

Roles narrow writes, not reads

Owner, admin, editor and member can change agents. A viewer can only read them. A few reads are role-restricted too. Only owners and admins can read the audit log.

Owners, admins, editors, members and viewers see every agent the workspace owns, unless their membership is limited to chosen agents. A role is not a mechanism for hiding agents; limit the member to chosen agents for that. Members with the Client role see only the client portal.

Actions restricted to administrators

These require admin because they create a route for data to leave the workspace:

  • Creating and deleting webhook endpoints.
  • Creating workspace integrations (stored credentials and a base address).

Members limited to chosen agents

A member can be limited to one or more agents. They see only those agents and nothing workspace-wide. Their role decides what they can do: a viewer reads the agents' settings, calls and results, and an owner, admin, editor or member can also change their settings.

What a limited member cannot reach

A member limited to chosen agents cannot archive, restore or delete those agents, change their address, turn on a public share link, or change their retention period. A limited member with an editing role can publish a new version, switch on website chat and voice, and create a website key; give them the viewer role if they should not.

Members still see an archived agent. A member with an editing role and access to the whole workspace can restore it.

How enforcement works

Access is checked the same way on every path. The editor, the in-app assistant, the MCP server and direct API calls all ask whether the user is a member of the workspace and whether their membership reaches this agent. See Security.

Agent settings are checked when they are saved and again before a version is published. Some are checked again during the call.

Publishable keys

A publishable key authenticates the website chat and voice widgets. It is scoped to one agent and carries an origin allowlist. Which surfaces a website may use, chat or voice, is set on the agent, and voice is off by default. The raw key is shown once at creation; only its prefix is displayed afterwards, and the stored form is a hash.

The origin allowlist

Because a publishable key is embedded in a public web page, the origin allowlist is what stops another site's visitors using it.

Allowed origins   https://example.com
                  https://shop.example.com

Request from      https://someone-elses-site.example   refused

A browser on a site that is not on the list is refused before any billable work runs. A request that sends no origin, such as one from a server, is not checked against the list. Keep the allowlist tight, and use a separate key for staging and local work.

The MCP server

The MCP server accepts a token from a browser authorisation flow with proof key exchange and dynamic client registration, or a personal access token a user creates in Settings. Either maps to a real user, so every operation runs with exactly that user's permissions. Access cannot exceed what the user could do in the application.

Scopes provide a coarse filter above that (read, write, delete, workspace admin). Operations that delete something, publish a version, or change what live callers get straight away require an explicit confirmation argument. The in-app Builder asks before the same operations, even when it is set to apply changes without asking.

Audit log

The audit log records who did what and when. Recorded events:

  • API key creation and revocation.
  • Member addition, removal, role change, agent-access change and ownership transfer, and invitations sent through the MCP server.
  • Workspace rename.
  • Workspace integration updates and deletions, and connector changes.
  • Phone number changes.
  • Campaign creation and launch.
  • Discovery start and stop.
  • Billing changes.
  • Changes to retention, personal information policy and data residency.

Agent settings changed through the MCP server, the REST API, the vai CLI or Builder are recorded under the name of the operation performed, for example agent.publish_version, with one entry for each change. A change that had to be approved first also records who approved it, as confirmed_by:

confirmed_byApproved by
personYou, in Builder, the in-app assistant, or an AI assistant that asked you itself
modelAn AI assistant, sending confirm: true
api_clientYour code, sending confirm: true to the REST API
cli_flagThe vai CLI's --yes or --max-spend flag
cli_promptA y typed at the vai CLI's prompt

Entries cannot be altered

Entries cannot be changed once written. Timestamps come from the platform clock, not the writer. Reads are restricted to owners and admins, and the log can be exported. Entries are kept for two years by default, a period each workspace can change, and are deleted with the workspace.

Each successful agent change made through the MCP server is recorded. Refused attempts are recorded too; operations that fail for other reasons are not. Reads of call content through the MCP server are recorded as access entries.

Account and workspace removal

Deleting a workspace begins a seven-day recovery window. After it, teardown deletes the workspace's recordings, releases telephone numbers, removes the billing record, and cascades deletion to everything the workspace owned.

An account is deleted by a daily job once the request is seven days old. Completion is recorded in a separate internal record that survives the deletion it describes.

On this page