MCP servers are 100% alive. Why we only index 15.

The liveness number is a selection effect

When we seeded the MCP server category in June, we added 15 servers with enough independent evidence to clear the ranking bar. All 15 respond to a live endpoint probe tonight. The overall Ghost Index for all 1,394 indexed agents is 58.8% — 572 agents have endpoints that no longer answer.

The 100% MCP liveness rate is not evidence that MCP servers are better-engineered than other agents. It's evidence that the servers we chose to index were, by construction, the ones that looked operational when we looked.

The same signals that push a server into the index predict whether it's running: active GitHub commits, presence in multiple registries, meaningful star counts. A project that abandoned its MCP server six months ago doesn't have those signals. It doesn't appear in our index. It also doesn't appear in our liveness count — because we never added it.

This is the selection effect. It appears across every high-quality data product. CoinMarketCap tracks tokens with enough market activity to have reliable price data. Arena.ai ranks models with enough Arena votes to have statistically meaningful Bradley-Terry coefficients. We track MCP servers with enough independent corroboration to say something confident about them.

How we scored the 15

MCP server scoring uses five signals weighted by reliability:

A server needs at least two independent signals to appear in the ranked list. The 15 that made the initial seed all had GitHub repositories with meaningful star counts and at least one public registry listing. Several have five-registry presence — the official MCP registry, Smithery, mcp.so, Glama, and Continue.

That multi-registry coverage also explains why they're all alive: a server that a curator at Smithery decided was worth listing, that the Glama team added to their connectors directory, and that the MCP steering group accepted into the official registry has cleared three independent quality filters before we ever looked at it. Those filters select for ongoing maintenance.

The gap between 15 and the ecosystem

MCP launched in November 2024. By mid-2026, public registries track hundreds of servers — the official registry, Smithery, mcp.so, and Glama each maintain their own lists, and the numbers are growing weekly. The gap between our 15 and the full MCP ecosystem is large.

Most of that gap is projects that announced a server, listed it in one place, and then stopped active development. They're technically discoverable — but they don't have the multi-signal corroboration that makes a ranking statement defensible. A server with 80 GitHub stars, listed on one registry, and last committed to four months ago is not something we can confidently say is better or worse than another one. We can only say it exists.

That's a meaningful distinction. The purpose of the MCP server rankings is not to list everything. It's to surface the servers with enough independent evidence that a developer choosing between them can make a data-informed decision. Adding entries where the evidence is thin doesn't improve that surface — it adds noise.

What the coverage roadmap looks like

The B8 ingestion pipeline — a fetch-and-evidence-check runner we built in June — is the mechanism for expanding coverage as the ecosystem matures. It pulls from public MCP registries, runs the evidence check, and promotes servers that clear the bar into the ranked list.

That pipeline is running in dry-run mode. The output looks something like this: of the hundreds of servers now listed across public registries, a meaningful fraction will clear the two-signal minimum threshold once the pipeline runs a full pass. The ones that don't are either too new to have accumulated evidence, or have stopped receiving updates.

If you build or maintain an MCP server and want it in the index, the path is concrete:

Servers that hit those thresholds will appear in the next ingestion cycle. You can also submit your agent directly to flag it for review.

What the 100% number is actually useful for

The 100% liveness rate for indexed MCP servers tells you something real: if you pick any server from our MCP server rankings today, you can expect its endpoint to answer. That's the practical value of a curated, evidence-ranked list versus a raw registry dump.

The overall 58.8% Ghost Index means that if you picked a random agent from the full index, you'd have a 4-in-10 chance of hitting a dead endpoint. The evidence-ranking filter changes that odds profile significantly — which is the point of evidence-ranking.

But the 15 servers are not the MCP ecosystem. They're the highest-evidence slice of it. The ecosystem is larger and more varied — and as the B8 ingestion pipeline expands coverage, the liveness rate for the MCP category will almost certainly come down from 100%, because the servers entering the index from that pass will have lower signal density and correspondingly lower evidence of ongoing maintenance.

That's what a representative sample looks like. We'll publish coverage statistics as the category grows, using the same methodology we used to correct the Ghost Index when the original number turned out to be measuring the wrong thing.