Legal

AI policy

What the agents in BetterCMS are allowed to do, what they can never do, and which crawlers may read a site you publish.

Effective date
Needs an answer

Draft under review

This document describes behaviour the product implements today and has not been reviewed by a lawyer. Where a commitment would need a contract rather than a build, it is shown as a marked gap.

The two rules

An agent plan always needs a person to approve it. Nothing an agent writes publishes on its own.

Those two rules are the whole policy and everything below is how they are enforced. They are not settings, so there is no switch that turns them off and no plan that removes them.

How a run works

A run moves through planning, awaiting approval, applying, review and done. A person is in the loop at two of those five points.

The plan comes first

Before anything runs, the agent shows a plan card stating the goal, the steps, the boundaries it will work inside and what the work will cost, either as a credit estimate or as the name of your own provider key. The two buttons are Reject and Approve and run.

A task is either one of the eleven named skills or a prompt you typed. A free-form prompt routes to the closest skill, and when it cannot tell, it falls back to the read-only audit, because a read is the safest thing to get wrong.

Changes arrive as drafts

Applying a plan produces drafts. Review opens on the documents the agent touched, one card per document, each with a real checkbox and a change count, and each changed field shown as a word-level before and after. You can confirm all of it, discard all of it, or apply only the documents you ticked.

Undo removes what is still a draft and tells you how many changes it kept, because those were already published or edited by somebody. A Stop control halts a run that is already going, and the note it leaves says whether anything had been changed.

A refusal is reported, not hidden

When a page refuses a write, the run leaves a note naming that page and the reason. It does not apply part of the job and call it done.

The highest action is a publish request

The BetterCMS MCP server gives an outside assistant twenty tools across three scopes: read, write to drafts, and run. The scopes are granted separately and a tool works only if its own scope was granted.

Publishing is not one of the tools. The highest thing an external agent can do is request a publish, and the request lands with a person who decides. A write against a published entry is staged as a proposal rather than applied to live content.

Six access options exist when you connect a client: read content, write content drafts, media, components, schema and SEO. Read is locked on, because a connection without it is useless. Publishing is not on the list at all.

  • A connection starts pending and works only after a person authorizes it.
  • Checks run before every call: the grant, the scope, the workspace switches and the credit budget.
  • Every call writes an audit line against its grant.
  • Keys expire, the raw value is shown once, and a revoke takes effect immediately.
  • One workspace switch suspends every grant at once.

Instructions and skills

Instructions are how a workspace sets the agent boundaries in writing. A skill teaches the agent how to do a task. A rule is a boundary it always respects.

The library is workspace scoped, so an instruction written once is read by every run in that workspace. Each instruction has a master switch and a per-project opt-out, and the agent page states how many instructions it is following on the run you are about to approve.

Of the eleven shipped skills, four write, three propose a change set for review, and four are read only. An admin can switch any of them off, and a skill switched off reads Off in AI controls rather than quietly disappearing from the list.

Models and your content

The product shows three speed tiers, Lite, Balanced and Max. It does not show a model name, and a request above the workspace ceiling is clamped down to it rather than rejected.

A tier can be bound to your own provider key. The work then bills to that key and runs against your own account with that provider. The raw key is validated, masked and dropped, and removing it falls back to the managed model and says so on the page.

Which providers sit behind the managed tiers, where they process the content, and whether anything is used for training are contract facts rather than build facts. They are not written here on an assumption.

Needs an answer

Training statement

Content processed by the AI features [ Training statement ].

The sentence a buyer looks for. It has to match the provider contracts exactly, so it cannot be written from the marketing side. Say what is true, including any exception.

Needs an answer

Model providers

The managed tiers are served by [ Model providers ].

Name each provider and the processing region. These are sub-processors and belong in the privacy notice list too.

What is recorded

An agent run is recorded the same way a person edit is. Every change goes to the audit log, every version of a document is kept, and a single field can be restored on its own rather than rolling the whole document back.

Run history exports as a CSV: when a run happened, which agent ran it, the task, the prompt, the status, how many changes were applied, whether it was reverted, and the cost. The agent also holds a role, works inside that role, and appears in the log beside everyone else.

Crawlers and answer engines

A site you publish with BetterCMS decides which crawlers may read it. The robots file is real and editable, and the default allows GPTBot, ClaudeBot and PerplexityBot by name. Removing one is an edit to that file.

Two other surfaces exist for agents. Every page and entry has a markdown twin, and the canonical URL answers a markdown request directly. An llms.txt file is generated or written by hand, and llms-full.txt is off by default, because a large site can exceed an agent context window.

What this policy does not claim

  • No certification. BetterCMS holds no SOC 2, no ISO and no HIPAA certificate today.
  • No accuracy guarantee. An agent drafts, a person reviews, and what you publish is yours.
  • No claim about third-party engines, which is why the crawler section says so in full.
  • No audit of a model provider, which is why the providers are a gap above rather than a list.

You stay responsible for what you publish, including the parts an agent drafted. The review step is in the run for that reason.

Changes to this policy

The two rules at the top are not going to loosen. If anything else here changes we will update the effective date, and a change that gives an agent more reach than it has today will be announced to the workspace owners before it ships.