Administration

Security and administration

How Volobox controls access for people and agents alike: roles, approval of agent writes, the audit log, sign-in options and AI controls.

Volobox treats people and agents under one permission model: an agent reaches only what its memberships and the person it works for allow, and its writes are approved and recorded. This page is for administrators and security reviewers who want to know how access is controlled, what is logged, and which switches an organization holds. For the security overview and review process, see Trust.

Roles and permissions

Organization administrators manage users, roles and account policies. Each settings area has its own permission, so people see only the administration they are allowed to change.

Inside a project, members hold a role: Viewer, Collaborator, Manager or Admin. Groups and agents can be members too.

Note

Being a project member does not give an agent a repository. Repository access is granted separately.

Project visibility

  • A public project is visible to every member of the workspace, never to the internet and never to people from a Client.
  • A private project is visible only to its members and the viewers added to it. Its tasks and requests stay hidden everywhere else they could appear.

How agents are authorized

In chat, an agent works in one of three modes:

  • Ask mode, read only.
  • Plan mode, a plan you approve.
  • Agent mode, where every write shows an approval card and nothing runs until a person approves.

A person can create a standing rule that pre-approves matching actions, with an optional budget on how often it may do so.

Note

A reply to a customer is never covered by a standing rule. A person always sends it.

Each team agent has two authorization profiles, each with its own mode: Interactive use (chat, team mentions, form assist) and Autonomous (workflow runs, assigned work items).

  • User delegated (Interactive use only): the agent does only what the person working with it may do.
  • Manager escalation (either profile): a write the agent may not do goes to its named manager for approval.
  • Self-authorized (Autonomous only): the agent acts within its own memberships without approval cards, under hard safety limits on how many writes it makes. Past a limit, the write comes back for approval or is refused.

Audit log

The audit log is an organization-wide history of work items, projects, shared docs and files, agent configuration, agent actions, message edits and deletions, and security events. It lives under Reports, so auditors can read it without configuration rights.

It exports to CSV, and a machine-readable export API lets a SIEM collect it with an API client registered for that purpose. Sign-in activity is reported beside it. See API, MCP and webhooks.

The Audit log under Reports: filters for date range, category, event key, and people or agents, above a table of events with actor and subject, and Export CSV.

Sign-in

  • Microsoft Entra ID: members sign in with your organization's Entra accounts. On Volobox Cloud this runs through the Microsoft 365 connection, with nothing to register in Entra. Self-hosted deployments use their own app registration. See Set up Microsoft Entra ID sign-in.
  • Password sign-in can be turned off for members once they sign in with Microsoft. Customers and external collaborators keep signing in to the Portal with their email and password.
  • Two-factor authentication uses an authenticator app or an email code. An administrator can leave it optional, turn it off, or require it.
  • Passkeys sign a person in with their device's lock or a password manager.

Note

Bring other identity requirements to the security review.

A person's account security settings: two-factor authentication on, an authenticator app configured, and two passkeys.

Apps that call Volobox on people's behalf sign them in through OAuth. See Build an OAuth integration.

AI controls

The AI kill switch turns AI off for the whole workspace, or for one area: assistant chat, coding sessions and reviews, workflows, system utilities, or the app Builder.

Warning

The kill switch is a hard stop. Running work fails at its next model call.

Under AI settings, administrators choose the models and the default coding agent, write workspace guidance every agent follows, govern what assistants remember, and decide who may use each MCP server. Credentials Volobox keeps for integrations are stored encrypted.

Data retention

Administrators choose how long reporting history (AI usage facts and storage snapshots) is kept. This does not touch tasks, files, workflows or the audit log.

Each person can download and erase what their own assistant remembers about them, and removing an account deletes that memory.

Clients stay in the Portal

People you invite from a Client reach only the Portal. They are never workspace members: they are not in the Team directories, cannot be mentioned and do not see presence, and Spotlight, announcements and public projects are closed to them. An external collaborator's access can end on a date you choose.

Reports

Reports shows AI usage and cost by operation, person, agent and model, with sandbox compute separately, plus agent runs, coding sessions, workflows and storage. It never shows prompts, completions or secrets.