guide · for MCP builders
You shipped a working MCP server. Now nobody can find it. Here’s the shortlist of registries and directories that actually get servers discovered — ranked by real traffic, with free-vs-paid on each.
I’ve built an MCP server — ServiceGraph’s — and published it to the official registry. When I first shipped it, I hit the wall every builder hits: the server worked, the demo was clean, and the user count sat at one. Me. The protocol solves how a model calls your tool. It does nothing for how anyone finds it in the first place.
That’s a discovery problem, and the fix is the same as it’s always been for new software: get listed where the people (and now the agents) doing the looking already are. For MCP servers that means the official registry plus a short list of directories that pull real visitor traffic. The number that matters most here is monthly visits— how many people actually pass through a directory and might land on your server.
There’s a second, smaller win worth knowing about. Your listing usually links back to your project’s site, so a spot on a high-authority directory is also a backlink — a modest boost to your own site’s SEO, separate from how the MCP server itself gets found. That’s what domain rating (DR, 0–100) measures, so I used it too — as the secondary signal it is.
Below is the list I’d give a friend — after submitting my own server through these and seeing what each one is actually like — grouped into tiers so you spend effort where it pays, followed by a walkthrough of publishing to the official registry with mcp-publisher, and what builders are saying on Reddit about what actually moves the needle.
You don’t submit to fifteen places one by one. Two assets do most of the work — a GitHub repo and an entry in the official registry. Get those right and the crawl-fed directories list you automatically; the rest are quick forms.
| directory | monthly visits | DR | listing | tier |
|---|---|---|---|---|
O Official MCP Registryregistry.modelcontextprotocol.io | 54K | 90 | free | 1 |
S Smitherysmithery.ai | 446K | 75 | free | 1 |
G Glamaglama.ai | 105K | 72 | free + paid | 1 |
M mcp.somcp.so | 238K | 72 | free | 1 |
P PulseMCPpulsemcp.com | 277K | 68 | free | 1 |
M mcpservers.orgmcpservers.org | 504K | 67 | free + paid | 1 |
M MCP Marketmcpmarket.com | 1.4M | 52 | free | 2 |
C Creati.aicreati.ai | 416K | 51 | paid | 2 |
A Awesome Claudeawesomeclaude.ai | 187K | 39 | free | 2 |
A AllYourTechallyourtech.ai | — | 37 | free | 2 |
C ClawToolsclawtools.com | — | 25 | free + paid | 3 |
M MCP Repositorymcprepository.com | — | 20 | free | 3 |
M MCP.Directorymcp.directory | 134K | 16 | free | 3 |
C ClaudePluginHubclaudepluginhub.com | 168K | 13 | free | 3 |
M MCP Server Hubmcpserverhub.net | 6K | 2 | free | 3 |
A note on the numbers: monthly visits are SimilarWeb estimates and DR is Ahrefs’ estimate. Both are approximations from third-party tools, not guarantees — treat them as directional, not gospel.
When the official registry shipped, the reaction in the community was, more or less, finally:
“So finally the day has arrived. Official MCP is here — an open catalog and API designed to solve how MCP servers are discovered. It doesn’t host the actual server code; it stores metadata (server.json) that points to packages in npm, PyPI, and Docker Hub.”
That metadata-only design is exactly why it matters: it’s the source of truth that other directories sync from. Publishing here is one action that ripples outward. When I shipped mine I just followed the official quickstart — here’s what that looks like:
server.jsonThis manifest is what the registry stores and what most directories display. ServiceGraph’s server is remote (Streamable HTTP), so mine declares a remotes block:
{
"$schema": "https://static.modelcontextprotocol.io/schemas/2025-12-11/server.schema.json",
"name": "co.servicegraph.mcp/servicegraph",
"description": "Datasets for founders: directories, newsletters, and agencies, with metrics attached.",
"websiteUrl": "https://servicegraph.co",
"repository": { "url": "https://github.com/nostrband/servicegraph", "source": "github" },
"version": "0.1.1",
"remotes": [ { "type": "streamable-http", "url": "https://mcp.servicegraph.co" } ]
}If your server ships as an npm or PyPI package instead, swap remotes for a packages array pointing at the published package. The description is the single most-read field on the whole listing — spend real time on it.
Namespaces are proof of ownership. The simplest path for most people is a GitHub namespace — io.github.<you>/<server> — via OAuth:
mcp-publisher login githubI use a domain namespace (co.servicegraph.mcp), which proves ownership by serving a signed key at /.well-known/mcp-registry-auth and logging in with mcp-publisher login http --domain …. Unless you specifically want your own domain on the listing, GitHub is less fuss.
mcp-publisher validate
mcp-publisher publishThat’s the whole loop. To ship an update, bump version in server.json and run the two commands again. The login token is cached, so re-publishing is a one-liner. Full reference: the registry’s official quickstart.
Grouped the way I actually prioritized them for myself. Submit to everything that’s free; spend your real attention on Tier 1.
The registry everything syncs from, plus the directories with the most real visitor traffic — where servers actually get discovered.
The canonical, Anthropic-backed registry. Stores server.json metadata pointing at your npm / PyPI / Docker package — it doesn't host code.
My take: The one submission that's non-negotiable. Most other directories pull from it, so publishing here propagates downstream. The listing itself doesn't currently produce an indexable backlink — the value is the syncing, not SEO. Publish with mcp-publisher (walkthrough above).
The most-used MCP platform among builders — hosting, one-click install into clients, and a large searchable catalog.
My take: If you list one place beyond the official registry, list here. In the Reddit threads I read while writing this, Smithery and PulseMCP are the two names builders reach for by default. High authority, real traffic, real installs.
Self-described "superset of the official registry" — 10,000+ servers, each scanned, scored, and testable in your browser.
My take: There's no submission form — Glama crawls GitHub, so add the `mcp-server` topic to your repo and it gets picked up, then claim the listing on the site to manage it. The in-browser inspector is a real differentiator: people can try your server before installing. Free listing is enough; the paid tiers buy placement, not access.
"The largest collection of MCP servers," with a clean submission flow.
My take: High authority and one of the fastest submissions on this list. A five-minute form for a DR-72 backlink and real browse traffic — easy yes.
submit →"Keep up-to-date with MCP" — servers, clients, use cases, and news.
My take: Heads-up: there's nothing to submit here directly — PulseMCP pulls from the official registry, so once you've published there (above) you'll show up automatically. It's the other name builders reach for alongside Smithery: high authority, strong traffic.
The "Awesome MCP Servers" list turned into a searchable site.
My take: The strongest organic-search footprint of the group — people land here from Google searching for a capability, not a server name. That's discovery traffic, the kind that converts. Free + paid.
submit →Real reach or a solid niche fit. Free or cheap, quick to submit.
Consumer-discovery framing — servers that connect Claude and Cursor to tools like Figma, Databricks, and Storybook.
My take: The biggest raw visit count here. The angle is end-users browsing for capabilities, so a clear, benefit-led description matters more than technical detail.
A broad "27,000+ AI tools, agents & MCPs" directory.
My take: Real organic reach, but listing is paid only — from a $69 one-time fee. You're also one category among many, so weigh the link and long-tail traffic against the cost; don't expect MCP-specific buyers.
submit →Curated directory of Claude tools, SDKs, MCP servers, and integrations.
My take: Worth knowing before you spend time: you submit by opening a pull request, and the maintainer seems to ignore most of them — getting listed is hit-or-miss. If you do land it, the curation means the listing carries more signal than the auto-aggregators, and it's a fit for Claude-targeted servers.
AI tools, MCP servers, APIs & agent skills, aimed at agent developers.
My take: Low traffic today but a clean free submission and an audience of builders rather than end-users. Fine as a fill-out-the-set listing.
submit →Smaller traffic today, but fast to submit — and uncrowded.
Community directory of tools, skills, plugins, and MCP servers.
My take: Young and low-traffic, but the free listing is fast to grab (there's also a $19 one-time paid listing if you want it). Uncrowded today and occasionally compounds.
submit →Search and discover official and community MCP servers.
My take: Thin authority, but a free listing costs you minutes. Submit and move on.
"3,000+ MCP servers & skills" with one-click install for Cursor, VS Code, and Claude Desktop.
My take: The one-click-install angle is nice for users who find you. Authority is low for now; treat it as discovery, not link juice.
Community directory of Claude Code extensions — commands, agents, skills, hooks, and MCP servers — with trust signals.
My take: Surprisingly high visits for its authority, and Claude-Code-specific. Listing is free — the paid options on the site are for visitors buying products, not for getting listed — so it's an easy submit if your server fits the Claude Code crowd.
"Best MCP Servers & Clients" directory.
My take: Lowest authority on the list. Free, so submit if you're being completist, but expect nothing.
submit →Several directories require a public repo, and some (like Glama) crawl GitHub directly and seed your whole listing from the README. No repo, no listing — so create one even if your server runs as a remote endpoint.
It's the source of truth the aggregators sync from. Do it before you start filling out forms elsewhere — some directories will pick you up automatically once you're in it.
Lead with the capability and the use case (“gives Claude access to X”), not the architecture. It's the one field everyone reads and most listings show it verbatim.
A clear install snippet and a one-line “what this is for” up top. Directories and users both judge you on the first screenful — and the crawlers literally lift it into your listing.
I spent a lot of time reading Reddit putting this together, and the takeaway is simple: the discovery problem is real — agents and humans alike struggle to find good servers, and the directories with actual visitor traffic are where servers get found. Here are a few of the threads I learned the most from — worth reading yourself:
“I maintain MCP servers that get 17,000+ calls/mo, and almost all the traffic has come from MCP registries and directories.”
“Or do you just publish your MCP on directories like Smithery and PulseMCP, then hope someone would notice it? … Not much user traction as of currently.”
“I kept watching Claude hallucinate MCP package names when I asked “is there an MCP server for X?” — the training data is stale and the model has no way to check.”
“I shipped a financial-data MCP server in 3 weeks — the Official MCP Registry, PulseMCP, mcp.so, Glama, Smithery, awesome-mcp-servers. Real things broke. Writing this up because I couldn't find this post when I started.”
The official registry is the one mandatory submission — it's the canonical source that most other directories sync from. The registry listing itself doesn't currently produce an indexable backlink for SEO, and the registry has low browse traffic on its own; the discovery and installs happen on Smithery, mcp.so, mcpservers.org and friends. Publish to the registry first, then list on the directories that actually send visitors.
Rarely, early on. Every directory in the top tier here offers a free listing that already gets you the backlink and the browse traffic. Paid tiers buy placement, not access. Spend the money only once a specific directory is already sending you users and you want more of them.
Yes, with the usual caveat: authority matters more than count. A link from a DR-90 registry or a DR-70+ directory moves your domain; a link from a brand-new DR-2 hub does little. That's why the list is tiered by domain rating — submit everywhere that's free, but spend your real effort up top.
The official registry is near-instant once mcp-publisher validates and publishes. Auto-aggregators that scan npm/GitHub can pick you up within days without any action. Manual-submission directories vary from minutes (auto-approve) to a week or two (human review).
Your server.json description and your README. Both the registry and most directories surface that text verbatim, and it's what a human decides on in two seconds. Lead with the capability and the use case, not the architecture.
Self-taught coder. Handled millions of fingerprints for several national governments. Ran a multi-petabyte web crawler. Arbitrage-traded 7 figures on Binance. Now working on ServiceGraph.