There is no single centralized MCP server registry in 2026. Instead, four complementary sources cover the discovery problem, and engineers use whichever fits the moment.
**`github.com/modelcontextprotocol/servers`** is the official reference repo maintained by the MCP team. It hosts canonical server implementations (filesystem, Git, Postgres, Slack, Memory, Brave Search, and others) and curates a README list of well-known community servers. Start here for trusted, vetted servers.
**mcp.so** is a community-curated marketplace organized by category (databases, devtools, communication, AI), with screenshots, install instructions, and reviews. Good for browsing what exists across the ecosystem.
**glama.ai/mcp/servers** is a searchable index that tracks usage statistics and version freshness, making it easy to separate actively maintained servers from abandoned weekend projects.
**smithery.ai** is an installer-style registry. Pick a server, click install, and it edits your `claude_desktop_config.json` or `.cursor/mcp.json` for you. This is the one-click experience that removes the JSON editing step.
None of these hosts the actual server artifacts. The MCP server itself typically lives as an npm package, a PyPI package, or a Docker image. The discovery sites point at those underlying artifacts.
MCP server discovery in 2026 is intentionally federated. There is no centralized registry analogous to npm or PyPI specifically for MCP servers. Four sources together cover the discovery need.
**`github.com/modelcontextprotocol/servers`** is the official reference repo maintained by the MCP team. It hosts canonical server implementations (filesystem, Git, Postgres, Slack, Memory, Brave Search, Time, Fetch) and curates a README list of well-known community servers. This is the trusted starting point and the source most install guides reference.
**mcp.so** is a community-curated marketplace. Browse by category (databases, devtools, productivity, AI), see install snippets, screenshots, and reviews. Best for discovering servers you did not know existed.
**glama.ai/mcp/servers** is a searchable index emphasizing operational metadata: usage statistics, version freshness, last commit dates. The practical value is filtering for maintained servers in a long tail where many entries have not been updated in months.
**smithery.ai** is an installer plus directory. Browse, pick a server, choose your host (Claude Desktop or Cursor), click install, and Smithery edits the right config file on your behalf. Removes the JSON-editing friction for non-technical users.
Two structural properties matter. First, these sites aggregate; they do not host artifacts. The actual packages live on npm, PyPI, or Docker registries. Second, they all describe servers built against the same protocol, so a server found in any source works in any compliant host (Claude Desktop, Cursor, Zed, Continue, Cline).
Quick reference
TL;DR
MCP server discovery in 2026 is federated: the official `modelcontextprotocol/servers` reference repo plus community indexes like mcp.so and glama.ai and the smithery.ai installer. No single centralized registry exists.
Imagine you want to find a specific kitchen gadget. There is no single website where every gadget lives. Instead, there is the official toolmaker's catalog (the reference gadgets they made themselves), a couple of independent shopping sites that list many brands and let you read reviews, and a helper that can drop a chosen gadget into your kitchen drawer with one click. MCP server discovery looks exactly like that in 2026. The official `modelcontextprotocol/servers` GitHub repo is the toolmaker's catalog. mcp.so and glama.ai are the shopping sites. smithery.ai is the one-click installer. None of them is the single true registry, and they all describe servers that use the same protocol, so anything you find in one place works in any MCP host.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Looking for the absolutely latest version of a server Discovery sites can lag behind. The npm, PyPI, or Docker registry holding the artifact is always the source of truth for version numbers.
Picking a server for production with hard maintenance and SLA requirements None of the four sources provides an SLA. For production use, vet the underlying repo's commit cadence and maintainer reputation independently.
Discovering vendor-hosted remote MCP servers Vendor-hosted servers (GitHub, Sentry, Linear, Stripe) are typically documented on the vendor's own product pages rather than in the general directories.
Quick scan · 2–4 min
Everything important, quickly.
3 min: no single registry in 2026, the four common sources (official repo, mcp.so, glama.ai, smithery), aggregators versus artifact hosts, cross-host portability as the unifying property, and the future direction (capability advertisement).
Real-world examples
Real products, models, and research that use this idea.
1An engineer looking for a Postgres MCP server checks `github.com/modelcontextprotocol/servers` first, finds the official reference implementation, and copies the config snippet into `claude_desktop_config.json`.
2A non-technical user installs the Linear MCP server via smithery.ai with one click. Smithery edits the Cursor config file behind the scenes and the user never touches JSON.
3An ops team picks a Sentry MCP server by comparing maintenance status on glama.ai/mcp/servers, choosing the one with recent commits and high usage over an abandoned alternative.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QWhy has the MCP ecosystem not converged on a single centralized registry like npm?
Sign in or go Pro to see the approach for this follow-up.
QHow does Smithery actually install a server into Claude Desktop or Cursor?
Sign in or go Pro to see the approach for this follow-up.
QWhat would a future official capability-advertisement spec change about discovery?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Expecting one canonical registry like npm for MCP servers. Discovery is intentionally federated across several sources; the actual artifacts live on npm, PyPI, or Docker, not on the discovery sites.