Skip to main content

The gates

Removing the badge is a separate add-on, and any plan can buy it — Starter included. It is not something a higher tier unlocks.That is why it resolves from a flag on the workspace rather than from the plan: remove_branding is the one entitlement a plan upgrade cannot grant by itself, so moving up a tier does not remove the badge on its own.Enterprise contracts include it at no additional cost — there is nothing extra to buy. It is still provisioned as the add-on, because the entitlement always resolves from the workspace flag and never from the tier.
Real-time human chat on hand-off needs Growth or above. On Starter a hand-off still works — it becomes a ticket rather than a live conversation.

Reaching the channel vs. acting through it

The MCP server is on every plan, Starter and trial included (agent_channel). Refusing the connection sold nothing — it just made the channel invisible until somebody paid to find out what it was.What Growth buys is the assistant acting: agent_actions is the gate on the agent executing writes rather than only showing and explaining.

Grants are capped per plan

A per-user MCP grant lets one end user point their own assistant at their own account. How many each user may hold depends on the plan: A trial, or any unrecognised plan, gets the Starter cap. Revoking a grant frees a slot.

Effective plan

Your workspace’s plan string is resolved to an entitlement tier before any check: Resolution fails closed. An unknown plan string never grants a feature by accident.

Active entitlement

A tier check is not the whole story. A workspace must also be entitled to serve at all, and three things independently revoke that:
  • the kill switch is on;
  • the trial has expired;
  • the subscription has lapsed.
Any of them and no gated feature resolves, whatever the plan says. The add-on entitlements are checked the same way — a purchased add-on on a workspace that is not entitled to serve grants nothing.

What a non-entitled workspace does

The behaviour is chosen per surface, so a lapse degrades rather than breaks.

The widget does not mount

The widget is told on load, before the cursor shows — no cursor, no crawl, no agent calls.

The agent goes quiet

The agent returns an empty, completed plan rather than an error, so a widget already on the page stops acting instead of throwing into it.

MCP stops answering

Both endpoints return “Agent channel not available for this workspace.”
Gated features stop entirely rather than degrading, since there is no partial version of most of them. Every gate is enforced on the server, not only in the dashboard, so an unentitled workspace cannot reach them by any route.

Credits, separately

Distinct from plan gating, a workspace consumes credits. When the pool is exhausted:
  • the widget does not start a new session;
  • a new conversation is refused before any model spend, while conversations already in progress continue;
  • narration stops and the widget runs text-only.
The widget is given its remaining narration budget on load and decrements it locally, so a session that runs out of voice mid-conversation goes quiet without a round trip per utterance.

Rate limits scale with the plan too

See Rate limits — Starter, Growth and Scale each get their own ceilings.