Every other registry indexes servers: a name, a URL and whatever prose the publisher wrote. The tool names and input schemas an agent actually has to match on exist in no index. So we ask each server directly, and store what it answers.
Measured, not claimed. This is the whole index by kind, 11,380 resources in total.
| MCP servers | 8,291 | |
| Agent skills | 2,388 | |
| Machine-readable docs | 264 | |
| Other callable resources | 180 | |
| OpenAPI services | 153 | |
| A2A agents | 63 | |
| Catalogues | 15 | |
| ARD registries | 12 | |
| Packages | 6 | |
| Plugin manifests | 3 | |
| GraphQL APIs | 2 | |
| Agent descriptors (ACP, OASF, AgentFacts) | 1 | |
| Datasets | 1 | |
| WebMCP browser tools | 1 |
Media types as publishers actually wrote them.
Filters match exactly, so a registry that does not normalise these silently drops thousands
of entries. We normalise all of them, plus both URN prefixes in circulation
(urn:air: and urn:ai:), on ingest.
application/mcp-server-card+json | 4,194 |
application/mcp-server | 3,706 |
application/ai-skill | 1,456 |
application/ai-skill+md | 789 |
application/mcp-server+json | 386 |
application/json | 144 |
text/markdown | 107 |
application/vnd.github.copilot-plugin | 102 |
application/openapi+json | 85 |
text/html | 84 |
application/a2a-agent-card+json | 60 |
application/vnd.oai.openapi+json | 51 |