Blog
hedwigaihedwigaihedwigai

MCP

Ten MCP registries and the use they can see: installs, pulls, connections and views

An MCP server reaches an agent through a registry: a list someone browses, or a marketplace inside the client that installs it in one click. Ten registries list between 199 and 119,868 servers. We read on 12 November 2026 what each one counts about the servers it lists, and the answer depends on where the registry sits. The two that install servers themselves, Cline and Docker, count installs. The large directories count stars or their own page views, or nothing at all.

62,550

servers in the official MCP Registry, two thirds of them served from a URL

1.47M

installs from Cline's marketplace across its 199 servers; the top 10 hold 52%

2.77M

pulls of Docker's Playwright image, the most pulled MCP server image

The ten, and what each counts

Size and visibility run opposite ways. Glama lists the most servers, and MCP Market draws the most search traffic of the sites that are only a directory. Cline’s marketplace lists 199 and Docker’s catalog 316, but they are the ones whose counts are installs, because their clients do the installing. Three of the ten could not be read at all: Glama and PulseMCP license their data only with a key, and LobeHub answers every automated request with 402, payment required.

Ten MCP registries, largest first: servers listed, what each counts per server, whether it can be read, and Google visits a month
RegistryServersCounts per serverReadable by a programGoogle visits a month
Glama119,868“Recent usage”, undefinedNo: its API needs a key and its licence forbids scraping and bulk republishing1,718 (glama.ai)
Official MCP Registry62,550NoneYes: an open API that asks to be read infrequently1,024
MCP Market53,663GitHub starsPages only29,977
PulseMCP21,556Estimated weekly visitors, a modelNo: a partner API, and its terms forbid automated access1,816
Smithery20,756Connections to each serverThe listing, about 260 servers; its search is failing2,401
mcp.so18,911Views of its own pagesEach server's page4,935
LobeHubNot readableNot readableNo: automated requests are answered 402, payment requiredShared with the rest of lobehub.com
GitHub MCP Registry770GitHub starsYes: an open APIShared with github.com
Docker MCP Catalog316Image pullsYes: Docker Hub's APIShared with hub.docker.com
Cline MCP Marketplace199Installs from inside ClineYes: one JSON answerShared with cline.bot

The official registry counts servers, not use

The official MCP Registry is what other registries copy from, and it keeps no usage at all. It returns no total either, so we counted every page, 626 of them. Two thirds of its 62,550 servers are remote, served from the maker’s own URL rather than installed. Among those that install, npm carries three times as many as PyPI. And 28,675 servers published their latest version in the last 30 days, which says more about how fast the list churns than about use.

The official MCP Registry, every page counted
Servers
Servers listed, latest version of each62,550
Active / deprecated61,569 / 981
Served from a URL (remote)42,197
Installable from npm16,482
From PyPI5,398
As a container image (OCI)1,769
As an MCP bundle (mcpb)1,692
Latest version published in the last 30 days28,675

Where installs are counted: Cline and Docker

Cline posts to its marketplace each time a user installs a server from inside the editor, so its count is the closest any registry comes to use. Installs gather at the top: 1.47M across 199 servers, 52% of them in the first ten, while the median server has 1,475 and 80 servers have fewer than 1,000. The reference servers that ship with MCP, Filesystem and Sequential Thinking, sit beside Context7, which feeds an agent current library docs.

Installs from Cline’s MCP Marketplace, top ten

Filesystem: 162,680Filesystem163kContext7: 135,084Context7135kBrowser Tools: 87,432Browser Tools87kSequential Thinking: 84,488Sequential Thinking84kGit: 58,369Git58kPlaywright *: 54,435Playwright *54kGitHub: 51,437GitHub51kFetch *: 49,845Fetch *50kFigma: 43,003Figma43kPuppeteer: 37,256Puppeteer37k

Docker publishes each catalog server as an image, and Docker Hub counts every pull since it was published. Pulls favour servers that run without a person at the keyboard: Playwright drives a browser, and Fetch, Time, Grafana and SonarQube suit pipelines, which can pull an image on every run. The catalog’s own signing image has been pulled 113M times, five times the 100 most-pulled server images together, which is why it is left out.

Pulls of Docker’s MCP server images since each was published, top ten

playwright: 2,774,039playwright2.77Mfetch: 1,986,900fetch1.99Msonarqube: 1,705,807sonarqube1.71Mgrafana: 1,179,264grafana1.18Mtime: 1,085,665time1.09Mduckduckgo: 1,055,322duckduckgo1.06Mcontext7: 666,686context7667kmemory: 577,650memory578kkubernetes: 467,960kubernetes468ksequentialthinking: 439,949sequentialthinking440k

Smithery counts something else

Smithery hosts servers and counts connections to each. Its most-used servers are not the ones Cline and Docker install: a gateway that puts many tools behind one connection leads with 419,019, followed by hosted servers for search, Gmail, Google Sheets, PubMed and company checks. Playwright, Filesystem and Context7 do not appear among the servers its listing reaches; GitHub has 5,908.

Two limits on this reading. Smithery’s search answered every query with an error on the day we read it, so a server could not be looked up by name, and its listing reaches about 260 of its 20,756 servers. The ranking below is of those 262.

Smithery's listing by connections, the 262 servers it reaches
ServerWhat it isConnections
pipeworx/gatewayA gateway to many tools behind one server419,019
henry-ships/sparkforgeSparkForge219,896
braveBrave Search87,579
bouch/uk-due-diligenceUK company checks60,881
adamamer20/paper-search-mcp-openaiPaper search59,106
gmailGmail57,738
googlesheetsGoogle Sheets56,138
theagenttimes/newsAgent News43,392
pubmedPubMed42,137
fenglucc/ko-financial-dataKorean financial data40,620

One server, five counts

Lining up the same eight servers across the five registries that publish a number shows how little the numbers agree. Playwright leads Docker by pulls and mcp.so by views, and is sixth on Cline. Filesystem leads Cline and is not in GitHub’s registry at all. Chrome DevTools has 55,143 GitHub stars and 2,820 views on mcp.so, and Cline does not list it. A star is attention, a view is a visit, a pull is a machine, and only an install from inside a client is a person choosing the server for their agent.

The same servers in each registry's own count; * marks a different maker's server under the same name (Cline's Playwright is ExecuteAutomation's, its Fetch zcaceres's)
ServerCline installsDocker pullsmcp.so viewsGitHub registry stars
Playwright54,435 *2,774,03934,08439,540
Filesystem162,680324,75414,311Not listed
Context7135,084666,6869,47964,664
Sequential Thinking84,488439,94913,385Not listed
Fetch49,845 *1,986,90013,169Not listed
GitHub51,437158,13834,70234,401
Brave Search33,936298,32110,101Not listed
Chrome DevToolsNot listedNot in the top 1002,82055,143

mcp.so also shows GitHub stars, but the reference servers all live in one repository, so Filesystem, Fetch and Brave Search each show the same 88,148. A registry that ranks by stars ranks repositories.

Run it yourself

Every count above came from the agentrank program in hedwigai’s shell, and the visits from traffic. Add mcp.hedwigai.com to your MCP client, turn the programs on once, and send these to its bash tool. The registry reads are free; traffic is billed per call.

$ programs enable agentrank
$ programs enable traffic
$ agentrank official all                 # counting: true at first; ask again in three minutes
$ agentrank cline 200
$ agentrank docker 100
$ agentrank smithery all --limit 100
$ agentrank github-mcp 100
$ agentrank mcpso playwright-mcp,filesystem,context7-mcp,sequentialthinking,fetch,github-mcp-server,brave-search,chrome-devtools-mcp
$ traffic estimate glama.ai,registry.modelcontextprotocol.io,mcpmarket.com,pulsemcp.com,smithery.ai,mcp.so

The program calls these toolkit routes. Their schema is at toolkit.hedwigai.com/agentrank/openapi.json.

Each command, its toolkit route, and what it answers
CommandRouteAnswers
agentrank official allGET /agentrank/official?query=allServers listed, active, deprecated, remote and packaged; the first call starts a three-minute count
agentrank cline 200GET /agentrank/cline?limit=200Every Cline server by installs, with the marketplace's total
agentrank docker 100GET /agentrank/docker?limit=100Docker's MCP images by pulls, catalog machinery apart
agentrank smithery all --limit 100GET /agentrank/smithery?query=allSmithery's reachable listing by uses, and how many it reached
agentrank mcpso <slugs>GET /agentrank/mcpso?slugs=…Views, installs and stars of each named server on mcp.so
agentrank github-mcp 100GET /agentrank/github-mcp?limit=100GitHub's MCP Registry by repository stars
traffic estimate <domains>POST /traffic/estimateEach registry site's estimated monthly Google visits

What these counts cannot say

None of them is a market share. Each registry counts only what passes through it: a server installed from npm or called at its maker’s URL never shows in Cline or Docker, and the official registry records nothing. Docker’s pulls run since each image was published, and Cline’s installs since each server was listed, so an older server has had longer to gather them.

The sizes of Glama, MCP Market and PulseMCP are what each states on its own page; we did not read their listings, because their terms or robots rules ask programs not to. Google visits are DataForSEO’s estimate of organic search traffic. They leave out people who arrive from inside a client, which is how Cline’s and Docker’s users arrive.

More head-to-head evals

GLM-5.3 against DeepSeek V4.1 Flash on 75 questionsTwo models answer the same 75 questions; one is always right, the other 14× cheaper.
MeasureGLM-5.3DeepSeek V4.1 Flash
Right, of 757571
Seconds a question, median3.91.8
Read the comparison
How we made a long-running AI agent for knowledge work 4× fasterOne cited knowledge work task through 13 experiments: 937 s to 233 s, with the model held constant.
MeasureFirst runBest since
Landscape, seconds937233
Memo, seconds759602
Read the report
Ten models plan a change to a spreadsheet: which do it in under a minute, traced call by callTen models drive one agent harness through the same spreadsheet request; six write the plan, timed and traced call by call.
MeasureQwen3 235BGLM-5.2Gemini 3.8 FlashClaude Opus 5
Seconds to the plan, median27485794
Plans written, of three runs2333
Read the comparison
SubQ, jev, Perplexity's Decider and Bespoke's Nimble on biotech diligence memos: finding the sentences a column is aboutFour decision models judge every sentence of ten biotech diligence memos for five columns, scored against sentences a person marked.
MeasureSubQjevDeciderNimble
Found, mean of five columns at section length64%65%66%68%
Mechanism of action, found at section length57%51%61%54%
Read the comparison
Leads with a reason to write: hedwigai ask, Apollo and treg on one brief40 leads each, every one checked: the right person, an inbox that accepts, and a dated reason to write.
MeasureaskApollotreg
Usable leads13 of 407 of 406 of 28
Cost per usable lead$0.037$0.27$0.23
Read the comparison
Self-hosted decisions for ask: Perplexity's open Decider in place of jev on 92 company questionsThe decisions inside hedwigai ask made by Perplexity's open Decider on two NVIDIA RTX 3090s, against jev.
Measureask with jevask with Decider
Questions right, of 928989
Wrong answers12
Read the comparison
Clinical decisions: jev and ask on System One, PubMedQA, SciFact and a 370-decision setEdgeEvals' System One board, PubMedQA, SciFact, and a clinical set built to System One's rules.
Measurejevask
Our 370 clinical decisions, right94.6%93.0%
PubMedQA, right (best EdgeEvals scored: 62.2%)77.0%74.4%
Read the comparison
Executive lookup for US-listed companies: Claude Opus 5.5 and ask on 95 questionsWho runs a public company, and how to email them: a frontier model from memory against hedwigai ask.
MeasureaskClaude Opus 5.5
Questions right, of 955963
Small companies right, of 20139
Read the comparison
Rental comps: floor-plan rents and amenities on 14 listings, five assistants comparedRent per floor plan and amenities for US rentals: hedwigai realestate-us against Claude, GPT, Gemini and Perplexity with web search.
Measurerealestate-usBest of four assistants
Floor-plan rents right, of 262623
Amenities claimed the listing doesn't state12129
Read the comparison
Company screening from SEC filings: jev, Vela 2.0 and ask on 92 yes/no questionsYes-or-no questions about public companies, answered by two decision models alone and by hedwigai ask.
Measureaskjev
Right, of the questions answered98%79%
Wrong answers, of 92219
Read the comparison
Defect localization in industrial images: four frontier models and a trained localizerFour frontier models asked to put a box around what they just described.
MeasureTrained localiserBest of four frontier models
Mean overlap with the defect (IoU, 0 to 1)0.4210.109
Read the comparison