How we check
Every number on this site comes from one of six sources, and every page says which one and when it was read.
1. The official MCP registry
We read every entry from the official MCP registry through its public API. Last read: 2026-11-18, 66,705 servers. Names, descriptions, versions, endpoints and packages are shown as the publisher wrote them.
2. The handshake check
For each active server that lists a Streamable HTTP endpoint, we open an MCP connection and ask for its tool list. That is all:
- We never call a tool, read a resource or send any user data.
- We try the 2026 protocol first and fall back to the 2025 handshake, the same way a current client does.
- At most one open connection per host, so a host that serves many entries is not flooded.
- Requests carry the user agent
agora-probe/0.1 (+https://openforallofus.com/probe). - A server gets 12 seconds. Failures for a network reason are checked again before they are counted.
Last check: 2026-11-18. Of 43,212 endpoints checked, 23,464 completed the handshake, 12,735 asked for sign-in, 172 asked for payment and 6,841 did not complete the handshake (no response, an HTTP error, or a reply that is not valid MCP). Another 435 rate-limited us and 379 list a templated URL; we count those as unknown and leave them out of the 43,212.
A server that answers the handshake is up. It is not thereby safe, correct or well behaved. Servers that run on your own machine (npm, PyPI, Docker packages) cannot be checked from outside and are marked so.
3. GitHub
For servers that link a GitHub repository we read stars, last push, licence and whether it is archived, through the GitHub API. Last read: 2026-11-18. 4,239 of 33,118 linked repositories did not resolve.
4. npm
For listed npm packages we read existence (200 or 404), the latest version and its deprecation flag from the npm registry, using its abbreviated metadata. We show the latest version's publish time or modified time when supplied; the abbreviated response may omit both. The npm downloads API supplies last-week downloads. Downloads include bots and CI and do not affect ranking. Package existence does not prove a server runs or is safe.
5. PyPI
The PyPI JSON API supplies existence, latest version, the latest release's upload time and whether all its files are yanked. Last-week downloads come from PyPI's public download table in Google BigQuery (bigquery-public-data.pypi.file_downloads), counted over the seven full UTC days before each refresh, as PyPI Stats asks bulk readers to do. When that query does not run, counts are unknown rather than zero. Downloads include bots, CI and mirrors. These signals do not prove safety, correctness or use by people, and do not affect ranking.
Packages read 2026-11-18: 16,863 of 16,863 listed npm packages and 5,405 of 5,405 listed PyPI packages checked. Counts use distinct package names; PyPI spelling aliases are normalized.
6. Signed reports
Anyone who runs a gate that signs its decisions can report that a server works, is broken or is unsafe, and attach the signed record. We verify the signature against the reporter's public key with the open @cedulon/core verifier. A signed report shows that gate allowed a call; today it does not prove which server answered, and the page says so. How to file one.
What we hide from search engines
Every server has a page, but we ask search engines to index only pages where we add something: a successful check or a live repository. Entries that look like tests, sit behind temporary tunnels, copy one template many times, come from a namespace with more than 500 entries, or link a missing repository are kept out of the index. Today 46,426 of 66,705 pages are indexable.
If you run a server
To stop the check against your endpoint, or to correct anything on your page, write to hello@openforallofus.com or open an issue on GitHub.