mcpbeat Sign in

Open Data

Everything we measure right now, as plain CSV. Rebuilt every night, last on 22 September 2026. Take it, chart it, publish it, sell what you build on it — CC BY 4.0, the only thing we ask is a link back. Month by month we also publish the totals. Day-by-day history is not published: it is the one thing here that cannot be reconstructed later, and it is what the API is for.

Downloads

free under CC BY 4.0 — a link back is required
servers.csv Every server in the MCP registry with its current state, uptime and real installs. 35 067 rows · 9.7 MB gzip, 2.3 MB →
skills.csv Agent Skills deduplicated by content hash, with token weight and how often each was copied. 61 913 rows · 26.0 MB gzip, 6.9 MB →
uptime-daily.csv Share of servers answering, by day. Starts the day we began measuring. 51 rows · 2 KB gzip, 831 B →
tools.csv Tools read from the servers themselves, with the date each appeared and disappeared. 269 846 rows · 149.3 MB gzip, 41.6 MB →
security.csv Security findings per server: what was found, how much, and what we actually checked. 35 067 rows · 4.3 MB gzip, 521 KB →

What is inside

measured by us, not copied from anywhere
Live status

Every remote endpoint gets a real MCP handshake every 15 minutes. The file records what it answered and when we asked — not what its README claims.

Real installs

Weekly downloads from npm and PyPI. Everyone else publishes GitHub stars, which say how many people liked a link, not how many run the thing.

Deduplicated skills

Agent Skills are matched by content hash, so one skill repackaged into forty repositories counts once — with the number of copies in its own column.

Caveats worth knowing

Read this before you build on the numbers.

Local servers have no uptime. Roughly half the registry ships as an npm or PyPI package that the client starts on your machine. There is no address to ping, so those rows carry installs but empty uptime columns — that is missing data, not zero availability.

“Not responding” is not always broken. A server behind a corporate network or one that requires an API key answers our anonymous handshake with a refusal. We record the refusal honestly; it usually means the server is running and simply will not talk to strangers.

Uptime history starts when we did. The daily file begins on the day we ran the first check, not on the day the ecosystem appeared. It grows by one row a day and cannot be backfilled — by us or by anyone else.

Using it

no key, no form, no rate limit
curl -O https://mcpbeat.com/static/data/servers.csv

Licence

CC BY 4.0

Attribution is required. The files are free for anything — a blog post, a dashboard, a paper, a commercial product, a model trained on them. The one condition is a visible credit to mcpbeat with a working link to the source. That link is the whole price: these numbers cost fifteen minutes of checks each, every day, and the link is what pays for the next day of them.

Republishing a chart, a table or a single number counts. Copy this line as is:

Source: mcpbeat — https://mcpbeat.com/data/

In HTML, an ordinary followed link is enough: <a href="https://mcpbeat.com/">mcpbeat</a>. What the licence does not allow is publishing the data as your own measurements, or stripping the credit and presenting the files as a dataset of unknown origin. Full terms: CC BY 4.0.

Daily snapshots

The same files, frozen each day. Current state is easy to copy — history is not: you had to be collecting it since August. Compressed only; tools are included on Mondays.

September 2026

2 files 17.7 MB

August 2026

21 files 257.2 MB