Vendor: github.com/modelcontextprotocol/servers · entry id 07-model-context-protocol-official-servers-repo · source 07-mcp-infrastructure.md line 184
The vendor ships and maintains the server itself. A wrapper built by Zapier, Composio or a similar third party does not count as official. The status was established by hand on 2026-08-24. On 2026-09-04 the recorded URL was a reachable repository or package: a server you install and run on your own machine over stdio. Callable after an install, not a remote endpoint.
mcp_status, verbatim from the file:
official
mcp_url, verbatim from the file:
https://github.com/modelcontextprotocol/servers
A tool below is one the server NAMES. Nobody has called it. That is the same two tier honesty rule the rest of the directory runs on: a named tool is research, and BENCH-TESTED stays the only claim that anybody ran anything.
141 of the 741 entries that record an official or community MCP server carry a harvested tool list. The other 600 are unmeasured, which is not the same as empty. Harvest last run 2026-09-27. Every name across every server is on the tools index.
No CLI found by the 2026-09-27 harvest across vendor docs, npm, PyPI, Homebrew and GitHub. That is a probe result, not proof of absence.
api_gate, verbatim from the file:
free (open source; dual-licensed Apache 2.0 for new contributions, MIT for existing code)
No documentation URL recorded.
727 of 1252 entries are in the same position. Blank is legal and it is published as blank.
Not measured. github_url, github_stars, github_last_commit and github_archived are null on every entry in this build.
The refresh rail specced in SPEC section 7.2 has not been run. An unstamped star count is a lie, so nothing is shown rather than something stale.
A github.com URL already appears somewhere in this entry, which is a seed for that rail and not a measurement of repo health:
github.com/modelcontextprotocol tied to the vendor by rule 1, the directory already classed this repo first-party and its owner is an Organization, confidence strong
The five most recently pushed, on 2026-09-08:
| Repository | Kind | Description | Stars | Pushed | Latest release |
|---|---|---|---|---|---|
| csharp-sdk | MCP server | The official C# SDK for Model Context Protocol servers and clients. Maintained in collaboration with Microsoft. | 4,521 | 2026-09-08 | v2.2.0 |
| ext-apps | SDK | Official repo for spec & SDK of MCP Apps protocol - standard for UIs embedded AI chatbots, served by MCP servers | 2,804 | 2026-09-08 | v1.7.5 |
| rust-sdk | MCP server | The official Rust SDK for the Model Context Protocol | 3,897 | 2026-09-08 | rmcp-v3.2.0 |
| kotlin-sdk | MCP server | The official Kotlin SDK for Model Context Protocol servers and clients. Maintained in collaboration with JetBrains | 1,450 | 2026-09-08 | 0.15.0 |
| java-sdk | MCP server | The official Java SDK for Model Context Protocol servers and clients. Maintained in collaboration with Spring AI | 3,685 | 2026-09-08 | v2.0.1 |
Kind is a heuristic guessed from the repository name, topics and description, not a fact the vendor stated. A repository count is activity, not a verdict. Read 2026-09-08.
No job tag on this entry.
Reason recorded by the tagging pass: no-job-fits. The official reference-servers repo. It is the spec's canonical reference, not a registry an agent queries and not a SaaS proxy.
981 of 1,252 entries are untagged. An empty list here means nobody has tagged this, not that the tool does nothing. The vocabulary is closed, so a tool whose job is genuinely not in it stays blank rather than being forced into the nearest tag.
1 source URL. Thin. The standing rule is at least two independent sources with the vendor's own site unable to be both of them, and this entry does not meet it. 100 entries are in the same state and they are listed on the methodology page. Raw sources field, verbatim:
https://github.com/modelcontextprotocol/servers
Every field above is rendered from directory.json exactly as the build produced it. Nothing is summarised and nothing is dropped. The one change made at render time is typographic and it is disclosed on the methodology page.