Finding MCP servers used to mean scrolling through GitHub repos and README files. Someone would publish a server, mention it in a Discord thread or a tweet, and maybe it showed up in a curated awesome-list. If you were building an agent and needed a Postgres connector, a payment gateway, or a web scraper, you searched GitHub and hoped for the best.

That worked when the ecosystem had a few hundred servers. It stopped working around the time the number crossed a thousand. AgentNDX exists because we hit that wall ourselves and figured other people were hitting it too.

What AgentNDX Actually Is

AgentNDX is a searchable index of MCP servers and agent skills. As of late June 2026, it tracks over 25,000 MCP servers and 95 agent skills across every major category — databases, payments, search, DevOps, cloud infrastructure, security, content, and more.

Each server listing includes the information a developer or agent needs to evaluate it quickly: what it does, how to install it, what transport it uses, whether it requires authentication, and what category it falls into. Skills get the same treatment — compatible agents, install commands, and what the skill actually handles.

The directory is available at agentndx.ai. There’s also an API for programmatic access, which is how other agents query the index directly.

Why a Registry Matters Now

The MCP ecosystem grew faster than anyone expected. Anthropic published the protocol spec, and within months hundreds of servers appeared. Then thousands. By mid-2026, the number passed 25,000 and kept climbing.

That growth created a discovery problem. If you’re building an agent that needs to interact with Stripe, there are multiple Stripe MCP servers. Which one actually works? Which one supports the auth model you need? Is it maintained? Without a central index, answering those questions meant reading source code and checking commit history on individual repos.

For AI agents, the problem is even sharper. An agent can’t browse GitHub. It needs structured, queryable data about available tools. That’s what a registry provides — not just a list, but metadata an agent can reason about when selecting which server to call.

How the Index Gets Built

The curation pipeline runs in batches. Each batch pulls from multiple sources: GitHub repositories, package registries, community submissions, and direct scraping of server announcement channels. Every server goes through a validation step before it enters the index — we check that the repo exists, the install command works, and the metadata is accurate.

The process is mostly automated but not fully autonomous. Edge cases still need human judgment: servers with ambiguous categories, abandoned repos that technically still run, forks that add meaningful features versus forks that just rename things. We run about one batch per day, each adding or updating a few dozen servers.

Skills curation follows the same pattern but at a slower pace. The skills ecosystem is younger and smaller, so the volume is lower. Each skill gets tested against its listed compatible agents before it enters the directory.

Architecture Decisions

A few choices shaped how AgentNDX works in practice.

We keep listings flat. Name, description, install command, transport type, auth method, category, tags. No deep code reviews, no scoring rubrics. That’s enough for a developer scanning the directory or an agent filtering programmatically, and going deeper would slow down curation without improving discovery.

Servers get assigned to one primary category — “database,” “payments,” “search,” “devops,” and so on. We started with about a dozen and expanded as the space grew. The taxonomy is deliberately opinionated: a server goes in one bucket even if it touches multiple domains. We argued about this early on. Turns out forcing a single category makes the index browsable, not just searchable.

Some servers get featured placements. These highlight well-maintained, widely-used servers that solve common problems. Yes, it’s a monetization layer — but it only works because the underlying index is trustworthy. A featured server that doesn’t deliver gets pulled.

What We Got Wrong

The first version of the index was too permissive. We indexed everything we could find — weekend experiments, unmaintained forks, projects that stopped working months ago. High numbers, low signal. Search for “postgres mcp server” and you’d get fifteen results. Twelve of them were abandoned.

We fixed this by adding a health check to the curation pipeline. Servers with no commits in 90 days get flagged. Servers with broken install commands get delisted. The index got smaller before it got bigger again, and the quality improved noticeably.

The other early mistake was ignoring skills entirely. When agent skills started appearing — capabilities delivered through CLAUDE.md files and similar config-based patterns rather than server processes — we treated them as out of scope. That was wrong. Skills and MCP servers solve overlapping problems, and developers need to see both options when evaluating tools for a workflow. Adding the skills directory in 2026 was overdue.

How It Compares

AgentNDX isn’t the only MCP directory. Smithery was one of the first. MCP.so has its own index. Anthropic maintains an official registry with a smaller, curated set. Glama offers another directory with its own curation approach.

The difference is scope and structure. AgentNDX indexes the broadest set — including servers and skills that other directories skip — while still running health checks and maintaining category organization. The API makes it queryable by agents, not just browsable by developers. And the skills directory is something most MCP-only registries don’t cover.

No single directory captures everything, though. Anthropic’s official registry carries authority a third-party index can’t match. Smithery has deeper integration with its hosting platform. We’re not trying to replace those — we’re covering the long tail they don’t reach.

Where This Goes

Protocols and servers are only useful if people can find them. That’s the boring, unglamorous work AgentNDX does — a searchable index with enough metadata to make decisions, updated daily, covering both MCP servers and agent skills.

If you’re building agents and burning time hunting for the right MCP server, start at the directory. If you’re building a server and want it discoverable, submit it. We’ll validate it and get it listed.

Browse the directory: agentndx.ai

FAQ

Is AgentNDX free to use? The directory and search are free. The API has a free tier and a Pro tier ($29/mo) for production agent workflows that need higher rate limits.

How often is the index updated? Daily.

Can I submit my own MCP server or skill? Yes. Same validation pipeline as everything else — if the repo exists and the install works, it gets listed.