Your MCP Server Is Listed, Not Callable: 55% Handshake Rate and 14% on the Current Protocol

Your MCP Server Is Listed, Not Callable: 55% Handshake Rate and 14% on the Current Protocol

Summary

  • Only 550 of 1,000 sampled remote MCP endpoints completed an unauthenticated handshake on 5 November 2026; 338 were correctly gated and 112 (11%) were broken.
  • Just 140 of 970 streamable-HTTP endpoints (14.4%) answered server/discover, while 290 servers required authentication without declaring it, so runtime behaviour must be verified rather than assumed from registry metadata.
  • Test with MCP Inspector on every deploy, declare authentication requirements accurately, and prune oversized tool lists to improve agent selection accuracy.
  • Synscribe helps B2B SaaS teams build agent-facing discoverability that converts, from MCP health checks to AI search visibility; speak with our team to close the gap between listed and callable.

On 5 November 2026, a census of 1,000 remote MCP endpoints drew its sample from the official MCP Registry and sent each endpoint an unauthenticated initialize request. The test ran twice, 46 minutes apart.

The results: 550 of 1,000 completed the handshake. 338 returned a credential challenge — correct behaviour for a protected server, not a failure. 112 failed outright. The stable headline is that roughly 55% of listed remote endpoints are callable without credentials, 34% are reachable but gated, and 11% are genuinely broken.

Most teams stop reading there. The finding that matters more is further down: only 140 of 970 streamable-HTTP endpoints, or 14.4%, answered the server/discover method introduced in the 2026-07-28 protocol revision. And 290 servers demanded authentication with no declaration of that requirement anywhere in their registry metadata.

A listing in the MCP Registry is a claim. A completed handshake is evidence. Those are different things, and the gap between them has measurable consequences for agent selection.

Listed vs. Callable: What the Numbers Actually Say

The registry is large. At the time of the sample it held 57,041 latest-version servers, with 38,270 declaring at least one remote endpoint. The November 2026 census drew 1,000 from that pool and probed each one.

The 338 authentication challenges are not failures. The MCP lifecycle spec is explicit: a server that responds to an initialize request with a credential requirement is behaving correctly. Those endpoints are live, speaking MCP, and waiting for the right caller. Treating them as broken — as some agent-readiness scanners do — misrepresents the server's state and, in practice, can cost a server points in a readiness score it does not deserve to lose.

The honest failure count is 112, about 11% of the sample. The breakdown:

  • Protocol errors. 42, including HTTP 404s.
  • Other HTTP errors. 31.
  • DNS and TLS failures. 29.
  • Timeouts. 10.

One further detail: 991 of 1,000 endpoints gave the same outcome when tested 46 minutes later. Nine changed state. That number is small in percentage terms, but it confirms that a single check at publish time is not sufficient. Server health is not static.

The two findings most teams miss

Protocol Lag

Of the 550 servers that completed the handshake, 423 (76.9%) spoke protocol version 2025-11-25. Eighty-eight fell back to 2025-06-18, 18 to 2025-03-26, and 18 to 2024-11-05.

The 2026-07-28 revision formally retires the stateful initialize / initialized exchange and replaces it with a stateless model using Mcp-Method and Mcp-Name headers. The adoption numbers show where that revision stands in practice: only 140 of 970 streamable-HTTP endpoints answered server/discover, the method that marks a server as current. 192 returned method not found — they speak MCP, but they have not updated.

The implication runs in both directions. Servers running older protocol versions are not yet callable by clients that implement only the July 2026 model. Clients that drop the legacy initialize handshake today lose reach to approximately 86% of public servers. Both sides of that equation are transition costs, not defects, but they are costs that show up in agent selection.

Metadata Integrity

The registry cannot validate whether a declared endpoint is live or whether its authentication behaviour matches its declaration. The official registry requirements cover namespace ownership and package verification. They do not probe remote behaviour.

The census makes the gap visible. Of the 911 servers that declared no required header in their registry metadata, 290 demanded authentication when called. Of the 89 that did declare a required header, 26 completed the handshake without one.

For an agent tool-picker that trusts registry metadata to determine authentication state, this means getting the caller's credentials decision wrong on a significant share of requests. The metadata is not a reliable proxy for runtime behaviour.

Why Callability Is a Selection Gate, Not Just an Ops Metric

Agent selection follows a sequence of four stages, and a server can fail at any of them.

  • Find. The agent locates candidates through discovery indexes.
  • Research. It evaluates freshness and declared capabilities.
  • Shortlist. It weighs finalists against the task.
  • Act. It integrates and calls.

A server that fails a handshake is filtered at the Research or Shortlist stage. One that misrepresents its authentication state adds friction at the Act stage. Both reduce selection probability.

Invisible to AI Agents? Synscribe helps B2B SaaS teams close the gap between listed and callable with agent-facing discoverability that converts.

Publishing to the registry gets a server to the door of the Shortlist stage. Answering a handshake, on a current protocol, with accurate metadata, is what gets it through. See Measure AI Visibility for how these stages connect to broader agent-facing discoverability.

Tool Surface Size

The census also captured tool counts. The median responding server lists 7 tools. 219 servers list between 1 and 5. Only 19 list more than 50, though the distribution has a long tail that includes at least one aggregator with 572.

This matters because of how agents select tools. A 2025 paper on retrieval-augmented generation for MCP (Gan and Sun, arXiv:2505.03275) measured tool-selection accuracy as a function of context size. Passing a model every tool definition in context produced 13.62% accuracy. A retrieval step that narrowed the set first produced 43.13%.

That is a 3.2x accuracy penalty for an oversized tool surface — a phenomenon the paper terms "MCP prompt bloat."

A server with hundreds of tools is not more capable from the agent's perspective. It is harder to use accurately. The MCP server distribution data gives broader context on how the registry population is shaped.

A Six-Step Health Check for Your MCP Server

The MCP Inspector CLI is the reference tool for this test. Run it against your published URL.

Step 1: Test the Unauthenticated Handshake

Send an initialize request offering protocol version 2025-11-25. A valid JSON-RPC result or a valid authentication challenge both mean the endpoint is live. Anything else is what agents are hitting.

npx @modelcontextprotocol/inspector \
  --cli \
  --transport http \
  --server-url <your-server-url> \
  --method initialize

Step 2: Test With Credentials and Confirm Your Tool List

Send the same request with your production credentials. Confirm the handshake completes and that tools/list returns your expected tools.

npx @modelcontextprotocol/inspector \
  --cli \
  --transport http \
  --server-url <your-server-url> \
  --header "Authorization: Bearer your-token" \
  --method tools/list \
  --format json

Step 3: Check for the Current Protocol

Test whether your server answers the server/discover method. A method not found response places your server in the 192-server group that has not updated to the July 2026 revision. If you are building a client rather than a server, implement a fallback to the legacy initialize handshake — dropping it today means dropping access to the majority of the public registry.

Step 4: Correct Your Registry Metadata

If your server requires authentication, declare it. If it does not, remove any stale header declarations. The 290 servers that demanded credentials with no declaration mislead every client that reads their metadata before calling.

Step 5: Prune Your Tool List

The median is 7. If you list significantly more, evaluate whether all of them belong in a single server surface or whether splitting by domain would serve agents better. The selection-accuracy data is clear: a smaller, well-defined set outperforms a large undifferentiated one.

Step 6: Automate the Test on Every Deploy

Nine endpoints changed state in under an hour. A passing test at publish time is a snapshot. Integrate the Inspector checks into your CI/CD pipeline. Use a timeout that reflects real-world agent patience — --connect-timeout 15000 is a reasonable starting point.

Quote the Rate With Its Denominator

The 55% figure will circulate. It will be quoted without context.

The correct form is: 550 of 1,000 sampled remote endpoints from the official MCP Registry completed an unauthenticated handshake on 5 November 2026. That phrasing carries the sample size, the population it was drawn from, and the date. All three matter, because the registry added 16,282 entries in October 2026 alone and the composition of that population is not fixed.

The same discipline applies to the 14.4% current-protocol figure: 140 of 970 streamable-HTTP endpoints answered server/discover in the same sample, on the same date. Strip the denominator and the number loses its meaning.

The Gap Between Listed and Callable

The MCP Registry is a directory. It records what servers claim about themselves. It does not verify the claims.

  • Liveness. Nothing checks that a declared endpoint responds.
  • Authentication. Nothing checks that runtime auth demands match the metadata.
  • Protocol version. Nothing records which revision a server actually speaks.

The census makes that gap concrete. Businesses that ship an MCP server and treat the registry listing as the finish line are leaving most of the work undone.

The endpoint needs to answer. The metadata needs to reflect what the endpoint actually does. The protocol version needs to keep pace with what agent clients implement.

For teams working on agentic visibility and agent-facing discoverability, the path forward is not more entries — the registry already has 57,041. It is callable entries, with accurate metadata, on a protocol clients can use.

Still Losing Agent Calls? Synscribe audits your MCP health, metadata accuracy, and AI search visibility so your server gets selected — not skipped.

To build an MCP presence that agents can actually reach and select, talk to Synscribe.

Frequently Asked Questions About MCP Server Health

What percentage of MCP servers are actually callable?

550 of 1,000 sampled remote endpoints from the official MCP Registry completed an unauthenticated handshake on 5 November 2026 — roughly 55%. Another 338 returned a valid credential challenge, meaning 34% were reachable but gated, and 112 (11%) failed outright. Always quote the sample size and date alongside the rate.

What does it mean for an MCP server to be callable?

A callable server is one an agent can reach and either use immediately or reach again with the right credentials. In the census, that meant a valid JSON-RPC result to initialize, or a valid authentication challenge. Anything else — 404s, DNS failures, timeouts, protocol errors — is what agents actually hit when they try to select you.

Is an authentication error a sign my MCP server is broken?

No. A server that answers initialize with a credential requirement is behaving correctly under the MCP lifecycle spec. In the census, 338 endpoints were gated this way and only 112 were genuinely broken. Treating a 401 as downtime misrepresents server state and can unfairly cost a server points in an agent-readiness score.

How big is the MCP Registry?

At the time of the sample it held 57,041 latest-version servers, of which 38,270 declared at least one remote endpoint. The census drew 1,000 from that pool. The population is not static either: the registry added 16,282 entries in October 2026 alone, so any figure you quote needs the date attached to it.

How do I check if my MCP server is healthy?

Send an unauthenticated initialize request offering protocol version 2025-11-25 using the MCP Inspector CLI. A JSON-RPC result or a credential challenge means the endpoint is live. Repeat with production credentials and confirm tools/list returns what you expect. Then add the check to CI so it runs on every deploy, not once.

Should my MCP server support server/discover or the initialize handshake?

Both, for now. Only 140 of 970 streamable-HTTP endpoints (14.4%) answered server/discover, so clients implementing only the July 2026 stateless model reach a small minority. Servers should adopt server/discover while keeping initialize; clients should implement initialize as a fallback or lose reach to roughly 86% of public servers.

How many tools should an MCP server expose?

The median responding server lists 7. Research on tool-selection accuracy found passing every definition in context scored 13.62%, versus 43.13% when retrieval narrowed the set first — a 3.2x penalty the paper calls MCP prompt bloat. If you list far more than the median, split the surface by domain.

How do I declare authentication requirements in the MCP Registry?

Declare any required header in your registry metadata, and remove declarations you no longer enforce. The registry cannot probe runtime behaviour: 290 of 911 servers that declared no required header demanded credentials anyway, while 26 of 89 that declared one completed the handshake without it. Mismatches break agent authentication decisions.

How often should I test my MCP server's health?

On every deploy, and ideally on a schedule. Nine of 1,000 endpoints changed state in the 46 minutes between the census's two runs — a small share that proves a publish-time check is only a snapshot. Wire the Inspector smoke tests into CI/CD with a realistic --connect-timeout 15000.

Does a listing in the MCP Registry guarantee agents will find my server?

No. A registry listing is a claim; a completed handshake is evidence. The registry verifies namespace ownership and package integrity, not whether a declared endpoint responds, matches its declared auth behaviour, or speaks a current protocol version. Answering a handshake with accurate metadata is what moves a server from Shortlist to Act.

Tags:
Published on November 19, 2026

Dominate ChatGPT and Google Search

Synscribe helps B2B companies with SEO & GEO using programmatic SEO approach. Book a call to find out how we help you win.