Limits

Limits

Written so you learn the ceiling by reading rather than by hitting it. Figures re-derived from source on 2026-08-24.

It creates nothing

No tool that creates a record, or that acts on many records at once, is available. That is a deliberate boundary, not a missing feature: creating and bulk-editing stay in the console, where you can see what is about to change.

Asked to create a segment or bulk-delete users, the correct answer is "use the console." There is no tool to approximate it with, and an assistant that claims otherwise is guessing.

No send_* tool exists among the 126. The server cannot send an email, an SMS, a push notification or a Slack message. It can generate their content, which writes a draft record into your project, but it cannot deliver anything.

It does delete, one record at a time, with no undo

Eleven delete tools are served. Each carries a destructive hint so a host can prompt before running it, but there is no undo on the Intempt side.

delete_journey_draft        delete_journey_transformer   delete_meeting_type
remove_experience_metric    delete_persona               delete_knowledge
delete_snippet              delete_brand_design          delete_studio
delete_character            delete_pose

delete_journey_draft is served over POST. A host that guesses at destructiveness by looking at the HTTP method alone will misclassify it as safe.

All eleven sit in the write wave, which is off by default and has never been called against production.

Collections are capped, and say so

Responses cap at 50 items and shed bulky per-item fields to fit a size budget the client enforces. Two keys tell you when:

A capped count is not a total. Read _truncated.total instead of counting what came back. A large enough raw response is refused by the client outright rather than truncated silently, so on a big project a list call can fail instead of returning a partial page.

Email addresses cannot be searched

list_users and list_segment_users both reject any search term containing @, which matches nothing. This is a bug on the Intempt side, not a usage error, and it is being tracked.

An empty result does not mean the user is absent. Ask for the user by name, or use a tool that resolves an address directly.

Three tools are currently failing

These fail on the Intempt side rather than in the server, so retrying does not help:

• get_meetings_summary
• list_snippets
• get_journey_list_count

They are named in the server's own instructions so an assistant does not burn a retry loop on them. Last re-measured 2026-08-11.

This list said six until 2026-08-11. Three entries turned out never to have been broken: they were platform-wide endpoints being called under a project-scoped path. A known-broken list is uniquely self-confirming, because it tells an agent not to call the very tools that would disprove it. Entries get re-measured, not inherited.

Two tools have no backend at all

search_project and list_active_workflows are registered nowhere. There is no cross-entity search backend and no workflows backend yet.

A few arguments are not wired yet

A small number of documented arguments are accepted but not yet passed through. The list only shrinks. If a tool appears to ignore something you passed, this is why. intempt registry describe <tool> in the CLI shows what a given tool accepts.

One project at a time

Every data call runs against one organization and one project. There is no cross-project query, and switching project changes what writes touch. The seven generate_* tools persist a real record into whichever project is active.