10,000 MCP Servers, One Fragmented Registry: The Discovery Gap Holding Back AI Tool Integration

The MCP Registry API returned 9,652 active server records on May 24, 2026. An independent census by Nerq, run four months earlier, indexed 17,468 servers from public sources. Both counts are accurate. The gap between them is the cost of building a protocol ecosystem without a standard discovery mechanism.

Method

Data for this post comes from the official MCP Registry API (queried May 2026), the Nerq Q1 2026 server census, GitHub's mcp-server topic count (May 2026), PolicyLayer's June 2026 security classification of 2,031 MCP servers and their 31,000 tools, and the SEP-1649 and SEP-2127 specification proposals in the MCP protocol repository. No figures were estimated.

The Count Depends on Who Is Counting

The MCP Registry API returned 9,652 latest-version records and 28,959 total versioned records on May 24. GitHub search counted 15,926 repositories tagged with the mcp-server topic on the same date. The Nerq Q1 2026 census swept multiple registries and repository hosts and found 17,468.

The discrepancy is structural. There is currently no standard mechanism by which a public MCP server can advertise its existence to a crawler or registry without a human first registering it or pointing an agent at it. The specification enhancement proposal SEP-2127 (superseding SEP-1649) defines server cards served at /.well-known/mcp/server-card.json, enabling pre-connection discovery. Major AI clients only began shipping support in April 2026.

An AI agent responding to a user prompt today cannot query which MCP servers on the web handle requests in a given domain. It is limited to its local configuration or curated registry — a directory model closer to pre-web Gopher lists than to modern DNS. Most of the servers indexed by Nerq but absent from the official registry are real, working servers that agents will never find autonomously.

Tool Counts Are Not Uniformly Distributed

PolicyLayer's June 2026 classification of 2,031 working servers reveals that the median MCP server exposes 8 tools, but the mean is 15.5 and the 99th percentile reaches 128. The largest single server in the dataset exposes 622 tools.

This skew matters for agents. MCP initializes all tools at connection time. An agent connecting to five mid-size servers at the 15-tool average level loads 75 tool definitions into context before it processes the user prompt. As server size scales past 30 to 50 tools, context pressure degrades task accuracy, particularly when tool descriptions are verbose. The current spec provides no mechanism for lazy loading or capability namespacing. A server that exposes 200 tools exposes all 200 on every connection.

The tool scale distribution also creates an indexing problem. Registries that ingest server metadata by establishing a live connection face wildly variable payload sizes — from 2-tool microservers to the 622-tool outlier. Without a static manifest format, registries cannot cache or rank servers without connecting, which means every registry refresh requires a live handshake per server.

The Protocol Is Growing Faster Than Its Signage

SDK downloads reached 97 million per month by March 2026, up from approximately 2 million at the November 2024 protocol launch — a 4,750 percent increase in 16 months. Q2 2026 closed with 9,400 published servers across the four major registries, sustaining a 58 percent quarter-over-quarter growth rate that held for three consecutive quarters.

The pace has created a large installed base built before features like server cards, streaming, and structured resource types existed. New servers are currently being created faster than the tooling to govern, index, or discover them matures.

An analogy: web crawlers discovered pages in the late 1990s by following links and guessing paths. Sitemaps (2005) and structured metadata standards (Schema.org, 2011) took years to gain widespread adoption. MCP is at the 1999 moment — abundant servers, minimal signage.

What This Means for Site Owners

If you operate an MCP server, discovery infrastructure matters as much as the tool inventory itself. A server absent from all major registries is invisible to agents that do not already know your endpoint URL. Registry submission is not automatic on publish; it requires an explicit step, and the Nerq Q1 2026 census found thousands of servers that had been published but never appeared in any curated registry.

Register in the official MCP Registry and at least one secondary registry. Major AI agent platforms maintain curated lists separate from the central registry. Check whether your server appears in both before assuming discoverability.

Implement /.well-known/mcp/server-card.json now, before it is mandated by the core spec. Two major AI clients began reading it in April 2026, and the SEP-2127 pull request targets a June 2026 merge. The schema is minimal: server name, version, transport URL, and tool summaries. A static JSON file served at the well-known path costs almost nothing to add. First movers gain passive discoverability in registries and crawlers that adopt the standard before competitors who wait.

On tool count: the median of 8 tools per server is not accidental. It sits in the range where agents comfortably hold all schemas in active context. If your server is approaching 30 or more tools, consider splitting by capability domain. The server card format in SEP-2127 allows flagging tool lists as dynamic, signaling to clients that they should query before assuming the full catalog — a partial mitigation until lazy-loading lands in the spec.

The registry fragmentation and discovery gap will close as the server card standard matures and adoption spreads. Site owners who invest in discoverability today — registry presence, well-known endpoint, right-sized tool inventory — position themselves ahead of the majority of the 9,600 plus servers that have not done so yet.