Autorules via MCP and the Adset AI assistant
Autorules can be configured not only manually in the builder but also through a conversation: via the in-cabinet assistant Adset AI or via an external MCP client (Claude, ChatGPT, Cursor) connected to the Adset MCP service. What autorules are, how they work and what guards they have is described in the article Autorules.
The idea is simple: you state the goal in words — for example, “create an autorule that turns off weak combinations in campaign X” — and the model itself looks at the campaign statistics, picks the metric, window and thresholds, describes the rule in words and creates it after your confirmation. The rule is always created in test mode (dry-run): it does not touch traffic until you separately ask to switch it to Live.
Who has access
- The autorules feature is enabled by the team plan (beta). If the team does not have it, the assistant and MCP will say so — contact support.
- The same permissions as in the cabinet apply: the team owner, tech and team lead create and edit rules, the buyer can only read them. Switching to Live is a separate permission; the buyer does not have it.
- The model sees only the campaigns and rules that you see.
How it works in Adset AI
Open the assistant in the cabinet and describe the task. The assistant then proceeds step by step:
- Clarifies the campaign. The campaign must have a stream set of at least two flows: the rule compares flows with each other.
- Looks at the statistics of the campaign by sets and flows over the last 7–14 days: clicks, installs, registrations, deposits, revenue, spend.
- Picks parameters from actual figures: which funnel stage has data (installs, registrations or deposits), what minimum data a flow needs to be evaluated, which window to use. Thresholds are not made up — they are derived from your campaign's volumes: on Facebook traffic this is tens to hundreds of clicks per flow, on popunder — thousands.
- Describes the rule in words: the campaign, the metric and period, exact thresholds with the figures that justify them, the action, the schedule. And asks for confirmation.
- Creates the rule in dry-run only after your “yes”. If there is an error in the configuration, the assistant fixes it on its own and retries.
- Shows the preview: what the rule would do with the campaign right now, which flows met the condition and which were held back by guards and why.
- Switches to Live only on a separate request. Say “switch to live” — the assistant will remind you that the rule will start changing real traffic, and then do it.
Example requests
- “Create an autorule that turns off weak combinations in the Sweet Bonanza FB campaign”.
- “Set up a rule: if a flow gives fewer than 60% of the deposits of the best one in the set, turn it off. Campaign — CI Sweet Bonanza”.
- “I want a notification when campaign X builds up volume but has no deposits”.
- “Show what my rules would do with campaign X right now”.
- “What autorules do I have and what mode are they in?”
- “Pause the ‘Worst Click 2 Install’ rule”.
What the model can do
Tool | Scope | What it does |
|---|---|---|
|
| The rule configuration schema, the list of metrics, limits and an example. The model fetches it once before composing a rule. |
|
| The list of rules and a full rule with its configuration and counters of admitted campaigns. |
|
| Preview: what the rules would do with the campaign right now. Applies nothing. |
|
| Candidate, admitted and hidden campaigns of a rule. |
|
| Creating a rule. Always in test mode. |
|
| Editing the name, description or configuration. Protected against concurrent edits: if the rule was changed in the cabinet, the model re-reads it. |
|
| Pause, resume, switching between dry-run and Live. |
|
| Admit, hide or restore a candidate campaign. |
A rule cannot be deleted via the assistant or MCP — only in the cabinet. A rule can be paused.
Safety rules
- Nothing is created without confirmation. The model first describes the rule in words and waits for your “yes”.
- A new rule is always dry-run. It calculates decisions on schedule and writes them to the log, but does not change traffic distribution.
- Live is a separate request. Switching to live mode happens only on an explicit command and only with the Live permission.
- Guards stay in place. The minimum of active flows, the daily change limit, and the pauses after changes and after manual weight edits work the same way as for rules from the builder.
Connecting an external client
For Claude, ChatGPT or Cursor to work with autorules, issue an MCP API key (see Adset MCP service) and enable the Read Autorules (autorules:read) and Manage Autorules (autorules:write) scopes for it. These scopes are disabled for new keys by default. A key grants no more than its owner has in the cabinet: role, plan and campaign visibility are checked on every call.
It is worth giving the external client the same sequence of steps that Adset AI uses: first get_autorule_schema, then the campaign statistics via query_stats grouped by campaign, set and rotation, then the rule, the preview, and only then Live.
FAQ
The assistant says autorules are unavailable. The feature is enabled by the team plan. Contact support.
The assistant refuses to create a rule. Check your role: a buyer can only view rules. Creating, editing and Live are available to the team owner, tech and team lead.
The rule is created, but nothing changes. That is expected: the rule is in test mode. Check the preview and the log, then ask to switch it to Live.
The assistant suggested a minimum data threshold that is too high or too low. Tell it — thresholds can be changed in words: “set a minimum of 300 clicks and a 7-day window”. Guidelines by traffic type are in the article Autorules.
Can it apply to the whole team at once? Yes: “a rule for all team campaigns”. Then all active team campaigns are admitted on saving, and new ones become candidates after the next cycle.