The AgentNDX directory just passed 1,000 curated MCP servers. Not scraped. Not auto-listed. Each one reviewed, categorized, and tracked for health.
A thousand is a useful number because it is big enough to show patterns. Here is what the data actually says about how the agentic ecosystem is taking shape.
Where the Servers Are
Category distribution tells the story of what developers are building agents to do.
| Category | Servers | Share |
|---|---|---|
| Data | 325 | 32% |
| Infrastructure | 188 | 18% |
| Productivity | 156 | 15% |
| Code | 101 | 10% |
| AI | 55 | 5% |
| Media | 52 | 5% |
| Payments | 42 | 4% |
| Web | 39 | 4% |
| Communication | 38 | 4% |
| Identity | 25 | 2% |
Data servers dominate. Nearly one in three MCP servers exists to connect agents to databases, data warehouses, APIs, and structured data sources. This makes sense. Before an agent can act, it needs to read. Data access is the foundation layer.
Infrastructure comes second. Cloud platforms, deployment tools, container management, DNS. The pattern here: developers want agents managing their stack, not just writing code.
The thin identity category (25 servers) is worth watching. Authentication is the hard problem in agentic workflows, and the fact that so few servers focus on it suggests the tooling is still catching up to the need.
How They Connect
Transport tells you how an MCP server talks to the agent.
| Transport | Servers | Share |
|---|---|---|
| stdio | 510 | 50% |
| HTTP (Streamable) | 507 | 50% |
| SSE | 4 | <1% |
The split between stdio and HTTP is almost exactly even. This was not the case six months ago. Early MCP servers overwhelmingly used stdio because it was simpler to implement. The shift toward HTTP reflects two things: remote deployment (you cannot run stdio over a network) and the MCP spec’s push toward Streamable HTTP as the recommended transport for production.
SSE is effectively dead. Only four servers still use it. If you are building a new MCP server today, pick stdio for local tools or Streamable HTTP for anything networked.
How They Authenticate
| Auth Type | Servers | Share |
|---|---|---|
| API Key | 631 | 62% |
| OAuth | 236 | 23% |
| None | 152 | 15% |
| x402 | 2 | <1% |
API keys are the default. Almost two-thirds of servers use them. This is the simplest auth model: paste a key into your MCP config, and the agent is connected. It works, but it has limits. API keys do not scope well, do not rotate easily, and give agents the same permissions as the human who created the key.
OAuth is the next tier at 23%. These are mostly platform integrations (Slack, Google, Salesforce) where scoped permissions matter. OAuth is harder to set up but gives you granular control over what an agent can access.
The 152 servers with no authentication are mostly open tools: file readers, format converters, local utilities. Nothing that touches sensitive data.
x402 micropayments remain a rounding error. Only two servers in the directory use x402 as their auth mechanism. Five total support x402 in any capacity. The protocol is live, but adoption is still in the earliest stage.
Verification and Health
Of 1,021 servers in the directory, 341 are verified. Verified means we have confirmed the server installs, connects, and responds to tool calls. That is a 33% verification rate.
This is not a quality problem with the other 67%. Verification takes time, and many servers require paid API keys or platform accounts to test. The unverified servers are listed because they come from credible sources (official platform repos, well-maintained open-source projects), but we have not personally run them.
Health tracking tells a more interesting story. Active servers respond to health checks. Inactive servers have gone dark. Unknown servers have not been checked recently. The active rate across verified servers is high, but the long tail of community-built servers includes plenty that were published once and never updated.
What the Numbers Do Not Show
Raw counts miss context. A few things the data cannot tell you on its own:
Quality varies wildly. A server that exposes two well-designed tools for a specific workflow can be more useful than one that wraps an entire API surface with minimal thought about how agents actually call it.
Categories overlap. A server like Supabase MCP shows up under data, but it also handles auth, storage, and realtime subscriptions. Category counts are useful for broad patterns, not precise taxonomy.
Usage is invisible. We can count servers. We cannot count how many agents are actually using them. A server with 10,000 daily agent connections and one with zero look the same in the directory. Usage data would change the picture significantly.
Five Takeaways
1. Data access is the killer app. Agents that read structured data and act on it are the dominant use case. If you are building an MCP server, connecting to a data source is the safest bet.
2. The stdio-to-HTTP transition is real. The 50/50 split will not hold. HTTP will win for production deployments because it works over networks. Stdio will persist for local development tools.
3. Auth needs work. API keys work for getting started. They do not work for production agent systems that need scoped permissions, rotation, and audit trails. The ecosystem needs better auth primitives.
4. x402 is not ready. Five servers out of a thousand. The protocol works. The adoption does not. This will either change fast when a major platform ships x402 support, or it will stay niche.
5. The long tail matters. Most of the 1,000 servers will never be widely used. But the long tail is where experimentation happens. Today’s niche server for geospatial data or medical records could become a category leader when the right use case hits.
FAQ
Q: How does AgentNDX decide which servers to include? A: Every server is reviewed before listing. We check that the repository exists, the install instructions work, and the server description accurately reflects what it does. We do not list servers that are forks with no modifications, abandoned projects with no commits in 6+ months, or wrappers that add no value over the underlying API.
Q: How often is the directory updated? A: New servers are added in weekly batches. Health checks run on a recurring schedule. The directory currently processes roughly 30-50 new servers per batch cycle.
Q: Where can I browse the full directory? A: The full directory is at agentndx.ai. You can filter by category, transport, auth type, and protocol. The AgentNDX MCP server also lets agents query the directory programmatically.