Agent access & safety

AI connections are powerful, so they are built to be safe by default. This page explains exactly what a connected assistant can and cannot do, and how to stay in control.

Scopes: read, propose, operate

Every connection has one scope, chosen when you mint a key or sign in:

| Scope | What it can do | | --- | --- | | Read | View data only — status, opportunities, coverage, evidence. No changes. | | Propose | Suggest changes that land in a Maark approval queue for a person to accept or reject. | | Operate | Make direct changes within the connection's project scope. |

Non-admin members get read by default. Only org admins can grant propose or operate.

Project scoping

A connection only reaches the specific projects you select. An assistant scoped to one project cannot see or touch another. There is no "all projects" backdoor.

Writes preview first

Write-capable actions default to a safe preview: the assistant shows what it would change before anything is committed. To actually write, it must explicitly confirm — so a stray request does not quietly mutate your data.

Keys are yours to control

• Keys (mk_…) are minted in Org Settings → External agent access and shown once.
• Revoke a key at any time to instantly cut off that connection.
• Guided sign-ins (claude.ai, ChatGPT) can likewise be revoked, and expired sessions are pruned automatically.

Everything is audited

Actions taken over a connection are recorded, so you always have a trail of what ran, when, and under which scope.

What an assistant can never do

Some capabilities are deliberately kept off the AI connection entirely:

• Publishing to your site is never done over a connection — publishing always goes through the human review flow.
• Owner-only operations (the internal marketing, outreach, and audit tooling) are never exposed to customer connections.
• Deep diagnostic tools that are not yet part of the connection surface stay out of reach until they are explicitly added.

Turning it on for your org

A project must be enabled for external agent access before any connection can reach it. An org admin enables this per project in Org Settings → External agent access, alongside key management. If a project is not enabled, connections simply cannot see it.