Anclum 審核 8,900 個 MCP 伺服器及智能體技能，檢查連結與授權資料

There are dozens of directories indexing MCP servers and agent skills now. Almost all of them rank entries, and a fair number take money for placement. That makes two questions hard to answer from the listing itself:

Does this link go to the project, or to somebody's affiliate wrapper?
If I run this thing locally, what licence am I agreeing to?

I spent the last few weeks aggregating about 8,900 entries (MCP servers, agent skills, open-source AI tools) and auditing them instead of just listing them. What follows is what turned up, including the parts that are still bad.

Finding 1: link cleanliness is worse than expected

In this space a lot of listings point at a redirect or an affiliate URL rather than the project's own site. It is easy to miss because the link resolves fine - it just lands somewhere with a tracking parameter attached, or on a wrapper page that forwards you after setting a cookie.

We re-checked each tool link against its canonical destination and dropped the ones that only existed as a wrapper. The current count on roughly 1,600 AI-tool entries:

Check	Result
Links that go through a wrapper/redirect	0
Links carrying tracking parameters	0
Links to a page that no longer answers	tracked separately, not hidden

The method is written on the site rather than in a blog post nobody updates, because the number is only useful if you can re-run it.

Finding 2: licence data is almost never shown, and it matters

If someone is going to run an MCP server locally, the licence is not a detail. It decides whether they can use it at work, whether they can ship it inside a product, and whether AGPL obligations apply to their service.

Almost no directory shows it. We pulled it from each repository and display it as a fact line on the entry page, together with whether the repo has a SECURITY.md. Out of about 6,200 repositories with a resolvable licence:

MIT: 4,168
Apache-2.0: 776
AGPL: 139
GPL: 94
MPL: 22
BSD-3: 19
No licence stated or not machine-identifiable: ~794

That last bucket is the interesting one. Nearly 800 projects publish a server people are expected to run, with no explicit licence.

Finding 3: freshness is the field nobody shows

Directories usually show a rank or a star count, both of which age badly. What I actually want to know before opening a project is when it was last seen alive.

Every entry now carries how many sources listed it, when it was first seen and when it was last seen. If an entry has not been re-verified in months, that is visible rather than hidden behind an ordering.

What is still bad

Being honest about the limits is the only way the rest of it means anything:

Most of the catalogue is harvested, not hand-curated. Quality between entries varies more than I would like.
Around 40 entries still have no usable description. They are marked rather than filled with generated text.
Some repositories are archived or have moved. Status is sampled, not guaranteed.
Coverage of Chinese-language tools is better than English coverage of the same tools, which is an accident of where the sources are, not a judgement.

Why no paid placement

The moment a directory accepts money for position, every other signal in it becomes unverifiable. So there is no sponsored slot, no featured tier, and ordering is by things you can check: sources, stars, recency, licence.

If you want to poke holes in the method, the entry pages show enough to do it. The audit rules and the current gaps are written up on the site.

It is at https://anclum.com and there is a read-only MCP endpoint at https://anclum.com/mcp if you would rather query it from an agent than browse it.