# Agent registry

A metadata surface where agent or service records can be published and found, typically carrying a name, endpoint, authentication requirement, declared skills and schemas.

Source: https://aecon.ai/glossary/agent-registry
Updated: 2026-09-15

## In practice

A builder searches a registry for a capability that extracts structured records from public web pages, and retrieves a server URL, an input schema and an authentication method.

## The distinction

A registry entry is a lead, not an offer. It rarely carries a current price, this buyer's eligibility, an acceptance test or a remedy, and the organisation that publishes the listing may not be the one that operates the endpoint or receives the money.

## Six gates between a listing and a purchase

A capability becomes something an agent can act on only when each of these has its own evidence.

| Gate | The question |
| --- | --- |
| **Visible** | Does a record exist anywhere? |
| **Reachable** | Does the declared route resolve for this buyer? |
| **Eligible** | May this buyer use it, in this market, on this plan? |
| **Available** | Is this version current, and how stale is the check? |
| **Suitable** | Does the capability and the offer fit the job? |
| **Trusted** | Who stands behind the result when it fails? |

The gates are cumulative. A provider can be visible, reachable and eligible while still being unpriced, or unable to say what counts as delivery. The honest state is then "candidate with unknown offer", not "buyable".

## Every count needs a denominator

Directory totals are publisher counts. A figure like "more than 23,000 resources" records how many records a surface displayed at a moment, not how many suppliers were reachable by a given buyer, quoted, accepted or used twice.

The same discipline applies to downloads, calls, unique payers, reviews, health badges and featured placement. Ask for the numerator, denominator, period, population and observer before comparing two registries economically.

## Freshness is a local clock

Aggregators scrape on their own schedules, some disclaim timeliness, and providers deprecate versions while existing subscribers keep them. There is no universal freshness clock, so a buyer's own stale threshold — and the reason a candidate was rejected for being past it — belongs in the record.

## Sources & context

- [MCP Registry](https://modelcontextprotocol.io/registry/about) — Model Context Protocol, Preview, 2026. Manages namespaces and standardised server metadata, points to separately published packages or remote URLs, and does not host artifacts or provide a security and uptime assessment. Checked 2026-09-13.
- [A2A agent discovery](https://github.com/a2aproject/A2A/blob/main/docs/topics/agent-discovery.md) — A2A Project, 1.0. Describes Agent Cards publishing provider and skill metadata, with optional signing and an authenticated extended card. Checked 2026-09-13.

## Related reading

- [Model Context Protocol (MCP)](https://aecon.ai/glossary/model-context-protocol)
- [Agent2Agent (A2A)](https://aecon.ai/glossary/agent2agent-protocol)
- [Service](https://aecon.ai/glossary#service)
- [Commerce protocol](https://aecon.ai/glossary/commerce-protocol)