The anatomy of an agent tool catalog

Every agent platform carries a tool catalog, and the shape of that catalog quietly decides what the agent is good at. We build one at ModuleX, so we exported the registry this site's own catalog pages are generated from and looked at the distribution. As of July 2026 it holds 179 integrations exposing 1,784 actions. Here is what the shape says.

The long tail is the catalog

The five biggest categories are Productivity and Collaboration with 21 integrations, Developer Tools and Infrastructure with 19, Marketing and Advertising with 17, Sales with 15, and Communication with 14. Together that is under half the registry; the other half spreads across fifteen smaller categories, from Finance and Payments down to a single Education integration.

That long tail is not filler. An agent's value shows up when a workflow crosses from a big category into a small one: a lead-routing run that ends in a CRM write, a support triage that ends in an issue tracker. The narrow categories are where the chain usually terminates.

Two auth models carry almost everything

Of the 179 integrations, 91 accept an API key and 60 accept OAuth; bearer tokens (16), custom schemes (23), and ModuleX-managed keys (5) cover the rest, and some tools accept more than one. The practical consequence: an organization connecting even a modest stack will hold a mixed bag of long-lived keys and refreshable grants.

That mix is why we treat the credential vault as a product surface rather than a settings page. Keys and grants age differently, rotate differently, and leak differently; a catalog this shape forces the platform to make storage, scoping, and rotation first-class.

The median integration has 7 actions

The registry's action counts are heavily skewed. PostHog tops the table with 78 actions, followed by Vercel with 56, Stripe with 50, incident.io with 46, and Freshdesk with 45. But the median integration exposes just 7, and most sit well under 20.

We read that skew as a design lesson, not a gap. A model choosing between a handful of clearly named actions makes better calls than one scanning an exhaustive API mirror, so most integrations ship the few operations people actually automate and stop there. The wide tools earn their width: their domains genuinely contain that many distinct jobs.

Selection is a measurable failure mode, not a hunch. A paper accepted at the ICML 2026 Workshop on Failure Modes in Agentic AI benchmarked six frontier models across 580 tool-use scenarios and found the best configuration reached only 74.8 overall accuracy, with weaker models failing by selecting the wrong tool in up to a quarter of their failures. That is a benchmark rather than a measurement of production traffic — roughly a third of its scenarios are deliberately adversarial — but it locates the error where a catalog designer would expect it.

What this means if you are building for agents

Three takeaways from the shape of the data. First, cover categories before you deepen them; chains end in narrow categories more often than they start there. Second, budget for auth diversity from day one, because the catalog will force it. Third, publish tasks, not endpoints: the median useful integration is seven good verbs, and the agent is better for it.