The agent cannot do what you did not allow.

Every request is checked: whose it is, what may be read, what may change and who approved it.

  • Every action passes one gate

    Each tool call is checked for organisation, role, scope, risk, budget and approval before it runs. There is no path around it.

  • Read-only first

    New connections start read-only. The scan that maps a system uses read requests only and proposes actions; nothing writes until you set each one.

    Automatic scanBeing built
  • Secrets stay secret

    Keys are entered only in the authenticated portal, stored encrypted and never shown again after saving. They never appear in a prompt or a log.

  • Approvals and never-auto

    Money and irreversible actions never run on auto. They wait on an approval card that shows the exact change, and a person you choose decides.

  • A full audit log

    Every read, action, approval and rejection is written down with who, when and what changed.

  • Organisation isolation

    Every record, task, file and log carries your organisation’s ID, checked on the server against your membership. An ID in a URL grants nothing.

  • Shadow mode

    Actions are prepared and logged, and nothing is sent. Before launch your team reads every answer in a shadow week.

    Shadow switch in the portalBeing built
  • Network limits

    The model never browses or calls arbitrary addresses. Each connector reaches only the hosts approved for it. Network-level blocking of private addresses and redirects is being completed before general availability.

  • Untrusted input

    Messages, documents and system responses are treated as data, not instructions. Permissions are enforced by the gate, not by the model’s good behaviour.

What happens to one request

  1. Identify

    Who is asking, matched by channel and a second detail before any record is shown.

  2. Scope

    Only the systems and fields this action is allowed to read.

  3. Risk

    Read, reversible, irreversible or money. The last two never run on auto.

  4. Approve

    If the mode says approve, a person sees the exact change and decides.

  5. Act and verify

    The action runs once, then Delvra reads the result back before confirming.

  6. Log

    Every step is written to the log, with undo where the system allows it.

Stop it at any time

  • Pause the whole organisationIn pilot
  • Take over any conversationIn pilot
  • Pause one agent or connectionPlanned

What we do not claim

We do not hold SOC 2 or ISO certification today. We will not show badges we have not earned. Ask us for our current security documentation and data processing terms.

Request security information

Start with one workflow. We set it up and run it with you.

Tell us which messages eat your team’s day. In 30 minutes you will see it running on a demo store, what it would cost you and what stays in your hands.