What Claude's connector directory actually counts

If you list an MCP server in Claude's connector directory, you have probably wondered what decides where it shows up. Anthropic's docs answer part of that, and a public JSON feed of every listing's scores and positions answers more. We pulled the feed, fitted its numbers and checked them against the docs. The worked examples use our own listing: we build Tofu, a Community connector (disclosure at the end).

The feed needs no credentials and returns up to 5,000 listings per call, with a cursor for the next page. Each listing carries its copy (name, one-liner, description, tool names, categories), its connection details and these fields:

Anthropic's docs describe the count behind the developer portal's Directory rank card: “distinct Claude accounts that sent your server any MCP message in the last 30 days.” Any message counts, initialize and tools/list included, and so do accounts “whose connection attempt never finished authenticating.” That is wider than the portal's Used a tool card, which counts only accounts that called a tool.

The feed's popularity_score fits that count on a log scale:

popularity_score =
  round(1000 × log2(N))

Here N is the number of distinct Claude accounts that connected through the listing in the last 30 days.

For small listings, where the possible values are far apart, every remote server's score lands exactly on one of them. Our listing, Tofu, scored 3,585 on 6 November; 2^3.585 ≈ 12, so 12 accounts in 30 days. Each doubling of accounts adds 1,000 points, and an account counts once however many calls it makes.

rank is a different number: the order the feed returns by default. On 6 November our listing was 8,854th of 9,517, in the bottom 7%. Two patterns show up:

• The first places went to listings added that day or the day before, none with recorded usage yet. Newer listings sat higher on average.
• Past the new arrivals, rank tracks accounts in a short recent window. Where the trending formula below let us recover that count, its correlation with rank was −0.93 (Spearman; negative because rank 1 is the top). The 30-day count fit less well.

So your 30-day number doesn't tell you where you sit by default.

Search is the same endpoint with a q parameter. Results are grouped by where the word matched, name matches first, and sorted by rank inside each group. What we measured for our listing on 6 November:

• tofu: 4 results. Tofu came 4th, behind three other listings with “tofu” in the name.
• deploy: 161 results. Tofu came 154th, matched only through its tool names.
• website: 432 results, no Tofu, though Tofu's description uses the word.
• host: no Tofu.

Search appears to read the name, the one-liner and tool names. Tofu matched deploy through its tool names, and website returned no Tofu at all, though its description uses the word. Don't count on your description.

One caveat: searching “tofu” inside the Claude app on 3 November put Tofu first. The app may order some searches differently, and we haven't measured generic words there.

Anthropic's docs say the Trending tag “marks servers in the top 10 by recent growth in that same message-based count.” In the feed, the ten listings with trending: true are the top 10 by trending_score, and four of them were Community listings.

For a listing with almost no earlier use, the score follows one formula, and 120 listings match it to the unit:

trending_score =
  round(1000 × log2(a)
  × (a − 1))

Here a is the number of distinct accounts in a short recent window.

The window's length isn't public; the score refreshes daily. Other listings don't fit it, and we couldn't pin down what else goes in. The score grows faster than the count, so it rewards accounts that arrive together. At the two cutoffs on 6 November, a listing with little earlier use would need this many accounts inside one window:

We don't know whether the app shows Trending per category, so the second cutoff may not matter. A slow trickle won't reach either.

Anthropic's docs say “there's no separate application for the Verified label or for where in the directory a listing is placed. Anthropic decides both during review.” A new listing is Community by default, and Anthropic “may then escalate listings flagged as highly useful to Claude users to Verified review,” where “reviewers run a functional test of each tool.” The escalation “is assessed automatically.”

The feed shows that usage is neither necessary nor sufficient:

• Front had about 62,000 accounts in 30 days and was still Community. So were lemlist and Hostinger's connector.
• Ample Computer, which puts apps online from Claude as Tofu does, was Verified with about 15 accounts. Several small companies' listings were Verified with no recorded usage.

The one promotion we could confirm is Adspirer's. Its numbers, posted in GitHub issue anthropics/claude-ai-mcp#954 while it was still Community, showed about 3,400 accounts using its tools, and it was Verified 4 to 10 weeks after listing. What else goes into the choice isn't public.

Verified also brings two things outside the directory:

• A page on claude.com, at claude.com/marketplace/connectors/<slug>, which search engines can index. The sitemap at claude.com listed 916 of these and none for a Community listing. Ample Computer's page loads; ours returns 404.
• Chat suggestions. The docs say every directory entry is eligible for Suggested Connectors, Claude's in-chat recommendations. In practice Community listings are left out. Issue #954, a bug report about exactly this, is open with no Anthropic reply, and a third-party test on 7 October found 0 of 9 Community connectors and 7 of 7 Verified ones. We haven't run our own test.

One kind of suggestion is still open to Community listings: Anthropic's help article “How Claude suggests connected apps” covers apps a person has already connected. Every person who connects is one more whose chats can bring your connector up.

1. Link to your listing. Popularity and Trending count accounts that connect through the listing, and the docs say that includes accounts whose sign-in never finished. Put claude.ai/directory/<your-slug> on your site, docs, onboarding and emails. We don't know whether people who add your server by URL count.
2. Put search words where search looks: the name, the one-liner and tool names. Tool names were the only reason our listing matched “deploy”. Renaming a published listing requires re-review, per the docs; when we edited our one-liner, the change went live the same day.
3. Time your push. Brand-new listings held the top of the default order, and Trending rewards accounts arriving together. If you can, have a launch ready for the day your listing goes live, and concentrate a push into a day or two.
4. Don't plan around Verified. There's no application, and usage alone doesn't decide it. Reviewers test each tool, so keep every tool working for a test account. Community listings get no claude.com page, so your own docs page has to explain your connector to search engines and AI assistants.
5. Read the right numbers. Watch rank in the feed, not only the 30-day count. In the portal, the Error rate (30d) card is not the health badge: the docs define them differently, for example on whether sign-in rejections count as errors. No public document ties either one to rank or to Verified.

• Feed: public, no credentials, paged with a cursor.

  https://api.anthropic.com/api/directory/servers?limit=5000&visibility=commercial,gsuite,gsuite-google&verified_tier=anthropic,partner,community

• When: one snapshot at 02:55 UTC on 6 November 2026, plus searches the same day with q added. The in-app search was on 3 November.
• Other sources: Anthropic's docs at claude.com/docs, claude.com's sitemap and marketplace pages, and issue anthropics/claude-ai-mcp#954.
• Method: we fitted both formulas to the snapshot and correlated rank with the short-window counts the trending formula gave us.
• Limits: the feed isn't documented as a stable contract, and the field meanings are our reading. Every number is from one snapshot, and cutoffs move daily. Window lengths aren't public. Correlation is not causation: we saw what goes with a better position, not what produces it. We didn't test chat suggestions ourselves.

We build Tofu, which puts apps and websites online from a chat, and Tofu is a Community connector in Claude's directory. We used its numbers because they are ours to show. Better visibility for Community listings is in our interest.

Three people connected Tofu from Claude follows three accounts that connected it in October, and the guide for an app made with Claude and the chat documentation walk through connecting it.