MCP Server Directories Compared: Glama, Smithery, mcp.so, MCPFind, and More

Six MCP server directories now compete for the same developers, and they are not interchangeable. Glama catalogs nearly 50,000 servers and layers quality grades and an in-browser inspector on top. The official registry from Anthropic, GitHub, and Microsoft is an API most developers have never visited. Smithery hosts servers directly and ships a CLI, while mcp.so leans on community contributions. MCPFind, for its part, produces five-client install configs and 95 setup guides. Each one solves a different problem, so picking the wrong one adds friction to a workflow that should be fast. This comparison pulls data from each directory to map who does what well, and where each falls short.

Why Server Count Is a Misleading Benchmark

The stat everyone cites is total server count. Glama lists roughly 49,837 servers. mcp.so comes in around 22,924, and MCPFind indexes 13,941. MCP Surge sits at about 12,698 spread across 42-plus categories. From there the counts taper off: the official registry at roughly 9,652, mcpservers.org at about 9,445, and Smithery hosting around 6,000-plus. The awesome-mcp-servers GitHub list is the outlier. Maintained by punkpeye and carrying 89.9K stars, it holds only 700 to 1,000 curated entries.

These numbers measure different things. A large directory built from automated GitHub scraping catches abandoned repos, one-commit experiments, and servers whose underlying APIs no longer exist. Glama syncs from the awesome-mcp-servers list as one upstream source alongside its own crawls, so its high count reflects breadth of discovery rather than independently verified quality for every entry. A pruned directory trades headline count for accuracy. MCPFind runs automated checks to remove listings with dead GitHub links, meaning a smaller index where a higher percentage of listings actually install today.

The practical question is not which directory lists the most servers but how many of those servers would run if you tried to install them this afternoon.

The Discovery-to-Deployment Gap

Finding a server is the easy part. Actually running it in your specific AI client is where most directories leave you on your own.

A typical directory page gives you a server name, a description, a GitHub link, and a star count. That confirms the server exists, and not much else. Getting it running in Claude Desktop means assembling a JSON config block with the right command, args, and env keys. Cursor wants a slightly different format. VS Code with GitHub Copilot has its own syntax, and Windsurf differs again. So developers end up copy-pasting examples from docs, GitHub readmes, and forum posts, and getting the format wrong on the first try more often than not.

MCPFind addresses this directly at the listing level. Every server page carries ready-to-paste configuration blocks for Claude Desktop, Claude Code, Cursor, VS Code with GitHub Copilot, and Windsurf. The right format is sitting there, no spec reading required. The guides section extends that with 95 step-by-step setup articles covering individual server types and client-specific edge cases.

Smithery takes a different approach at the deployment layer. The smithery install CLI handles installation programmatically, and for servers that Smithery hosts directly, you receive a managed endpoint rather than a local process to configure. The Toolbox meta-MCP lets you access multiple Smithery-hosted servers through a single connection.

Glama solves the deployment gap at the other end with a hosted gateway. For supported clients, Claude Desktop, Cursor, and Claude Code, you connect to Glama's endpoint rather than running the server yourself. That eliminates local configuration entirely for servers inside Glama's network.

What the Official MCP Registry Actually Is

The official MCP registry at registry.modelcontextprotocol.io trips people up because the name sounds like a browsable directory. It isn't. The registry is a machine-readable API backed by Anthropic, GitHub, and Microsoft.

Developers query the registry programmatically to discover servers by namespace. Authentication uses reverse-DNS naming conventions and GitHub OIDC, which lets namespace owners prove they control a domain or GitHub organization before claiming a namespace. This prevents anyone from registering com.stripe.mcp or io.github.anthropic without proving that affiliation.

For AI clients and tooling that need to auto-discover servers in code, the registry provides the canonical source of truth. For a developer browsing to find a Notion integration to try this afternoon, it provides nothing without a client that surfaces its data. Think of it as the DNS layer for MCP servers: essential infrastructure, invisible in day-to-day use.

Quality and Trust Signals Across Directories

Every directory answers the question "how do I know this server is worth using?" differently.

Glama uses A through F quality grades based on criteria including documentation completeness, repository health, and tool schema quality. That grading layer, combined with an in-browser inspector that lets you test a server's tools before installing, gives Glama the strongest pre-install evaluation flow of any public directory.

The awesome-mcp-servers GitHub list earns trust differently: through star counts and the reputation of the list maintainer. With 89.9K stars on the list itself, inclusion carries social proof that a repository with 12 stars on its own cannot replicate. The tradeoff is that curation moves at human speed, so new quality servers lag before appearing.

The official registry's namespace authentication provides a different kind of trust signal. A server claiming the com.stripe.mcp namespace has had that claim verified against Stripe's actual domain ownership. That does not guarantee code quality, but it eliminates impersonation.

Smithery carries a 4.79 out of 5 star rating from user reviews, which reflects hosting quality and developer experience. A separate concern sits below that rating: a March 2026 security scan of sampled servers in Smithery's directory found that 22% had security findings. Smithery is the hosting platform rather than the author of the server code, but the stat points to a gap that no directory has fully closed. Community-sourced server code requires independent review regardless of the platform hosting it.

MCPFind's quality signal is operational rather than evaluative: live link pruning removes servers with dead repositories automatically. That confirms what currently exists rather than whether the code is good, but it reduces the time a developer spends clicking dead links in search results.

How Each Directory Compares

Curation vs Comprehensiveness

The tradeoff between broad indexing and careful curation runs through every directory decision.

Comprehensive indexes like Glama and mcp.so catch new servers faster and give developers a better chance of finding something niche. The cost is signal-to-noise: a search for a Snowflake integration returns results alongside unmaintained forks and servers with broken dependencies unless the directory has a grading or pruning layer on top.

Curated indexes like awesome-mcp-servers and mcpservers.org invert this. Their smaller counts reflect deliberate inclusion decisions. A server that clears the bar for the awesome-mcp-servers list has been looked at by a human. The tradeoff is coverage: specialized servers from smaller teams may never appear, and the review queue creates lag between a server shipping and being discoverable.

Live pruning sits between these approaches. Removing dead links reduces noise without requiring manual review for every entry, but it does not catch low-quality servers with functional repositories.

Neither model is objectively correct. If you are weighing ten candidate servers before picking one, a grading system like Glama's earns its keep. If you already know the server name and just need the install config, you want a directory that hands it over immediately.

AI Discoverability and Structured Data

An emerging concern for MCP directories is whether AI agents can discover them programmatically. Traditional search optimization targets crawlers. Generative Engine Optimization, GEO for short, targets the retrieval systems that AI tools use when answering questions or routing tool discovery.

Directories that expose structured data, llms.txt files, and machine-readable server metadata make themselves queryable by AI agents that want to find tools programmatically. The official registry is the most explicit example: it exists specifically to serve machine queries, not human browsers. For directories targeting developers who already use AI assistants as part of their research workflow, this layer is increasingly relevant to long-term discoverability.

Which Directory Fits Your Workflow

No single directory wins across all dimensions, and most developers end up using more than one.

Glama is the one to reach for on scale, grading depth, and that in-browser inspector. It shines when you want to judge a server's tool quality before committing to an installation, or when you want a hosted endpoint for Claude Desktop, Cursor, or Claude Code without running a local process. For surveying an unfamiliar category before you commit, it's the best starting point.

The official registry's strengths are authority and machine readability. Reach for it if you're building tooling that needs to discover servers programmatically, or if namespace verification matters for your security requirements. Just don't expect to browse it casually.

Managed hosting is Smithery's lane. If you want a server running in the cloud without provisioning infrastructure yourself, the smithery install path and the Toolbox meta-MCP cut setup time meaningfully. Given the security scan data, review the specific server's code on your own first.

For community trust, awesome-mcp-servers is hard to beat. An 89.9K-star list carries social proof that no automated index replicates. Treat it as a signal layer on top of a larger directory when you want to know which servers the community has actually endorsed.

MCPFind earns its place on the installation step, especially for developers working across multiple AI clients. The categories section organizes 13,941 servers across domains like devtools, databases, and productivity, and every listing carries the exact JSON config block for each of five clients. The guides cover server-specific setup with client-by-client instructions. If you're bouncing between Claude Desktop, Cursor, VS Code with Copilot, and Windsurf, having the right per-client format in one place removes a lookup step from every new server install.

Which directory serves you best comes down to where you spend your time in the workflow. Start at Glama for evaluation and scale. When you already know the server and just need it installed in a specific client, MCPFind is the faster path, and for programmatic tooling the official registry is the one to query. Checking community validation? That's awesome-mcp-servers. And when managed hosting is the goal, weigh Smithery against the server's own code.