Key findings
- We track 453 MCP servers. The median server has 1 GitHub star; only 27 (6%) have 1,000 or more.
- 40% had a commit in the last 90 days, while 65 (14%) have been quiet for over a year.
- 172 servers (38%) declare no license on GitHub.
- 86% show signs of writing to external systems and 85% ask for an API token or other credential.
- Of 453 servers in our benchmark, only 26 could be started automatically and answer the MCP handshake with no setup.
What the ecosystem is building
Servers are grouped into categories by what they connect an AI assistant to. Developer tooling dominates; everything else is a long tail.
| Category | Servers | Share |
|---|---|---|
| Developer Tools | 250 | 55% |
| Search & Retrieval | 70 | 15% |
| Monitoring & Observability | 58 | 13% |
| Databases | 16 | 4% |
| Productivity & Docs | 15 | 3% |
| Files & Storage | 14 | 3% |
| Web Scraping & Crawling | 9 | 2% |
| Browser Automation | 8 | 2% |
| Cloud & Infrastructure | 7 | 2% |
| Payments | 6 | 1% |
Popularity is highly concentrated
A small group of servers gets almost all the attention. 351 servers (77%) have fewer than 10 stars — many are personal experiments published to the official registry. The most-starred repositories we track:
- Filesystem MCP — 91.1K stars
- Context7 — 62.8K stars
- Playwright MCP — 38K stars
- GitHub MCP Server — 33.5K stars
- Agent Skills Search Server — 26K stars
Stars measure attention, not quality. A low-star server from a vendor you already use can be the right choice; see how to choose an MCP server.
Maintenance
110 servers (24%) were updated in the last 30 days and 179 (40%) in the last 90. The MCP specification and client apps still change often, so a server untouched for a year is more likely to break with newer clients. Track recent activity on the recently updated ranking.
Licensing
281 servers (62%) declare a license. MIT (234), Apache-2.0 (39), AGPL-3.0 (4) are the most common. Without a license, the code is not open source in the legal sense: you can read it, but reuse and modification are not granted. That matters if you plan to fork or bundle a server.
Permissions and risk signals
We scan each server's README and metadata for what it can do. Of 453 profiled servers:
- 438 use network access
- 389 can write to external systems (create, update, send, delete)
- 384 need an API token or other credential
- 172 access the local filesystem
- 87 connect to a database
Only 217 document security, permissions, or scopes in their README. Because most servers can change data somewhere, 397 land in our High risk level: most MCP servers exist to act on your behalf, which is useful and risky at the same time. Read MCP server security for how to judge this.
How many actually start?
We try to start each server in an isolated environment, with no real credentials, and ask it to list its tools, resources, and prompts. Results across 453 servers:
| Outcome | Servers |
|---|---|
| Not run: no usable start command in the docs | 276 |
| Not run: needs credentials (API keys or tokens) | 77 |
| Not run: runtime not yet supported, or command unsafe to run | 58 |
| Started and answered the MCP handshake | 26 |
| Started, but no handshake before the timeout | 16 |
Most were not run at all: either we found no start command we could run safely, or the server needs credentials we deliberately do not provide. Servers that did start expose a median of 9 tools; the largest exposes 111. See the most tools ranking and how we benchmark.
About this data
Server listings come from the official MCP registry plus curated additions and reviewed submissions. GitHub metrics are refreshed daily, and every number on this page is recalculated from the live dataset — it changes as the data does. Risk signals are inferred from documentation and can miss behaviour that is not documented.