The state of the MCP registry, October 2026

1. The registry in numbers

The registry lists an entry per server version. mcpcost takes the latest version of each server. It probes every entry with a remote URL, and every entry whose only package is on npm or PyPI. Entries with only an OCI or MCPB package are not probed.

By result

Three groups. Alive: the server answers initialize and tools/list. Gated: the server is reachable, but it wants auth, a payment, an environment variable or an argument first. Down: everything else.

2. What changed in October

The first full run was on 2026-10-03. The table compares it with the run of 2026-10-24. The rows cover remote servers only, so the two runs measure the same thing. Package servers are in the second table.

Token cost

Exact counts start on 2026-10-05, the first run with an API key. From 2026-10-15 the alive set also holds package servers. Their tool sets are smaller, so the median goes down as they join. The estimate row covers every run.

Run by run

25 daily runs. A point is missing where a run could not separate remote from package servers (2026-10-15 to 2026-10-20). Hover a point for its value, or open the table.

Alive remote servers

Alive, share of remote servers

Median tokens of a tool set in Claude Opus 5.5, alive servers, exact

Servers that need auth, share of remote servers

Protocol 2025-11-25, share of alive servers with a known version

Registry entries, latest versions

3. Token cost

A client sends the model the name, the description and the input schema of each tool. It does this in each request, also when the model calls no tool. The numbers below are the cost of the tool definitions of one server, as the Anthropic count_tokens endpoint counts them for Claude Opus 5.5. The endpoint is free and calls no model. Claude Code puts mcp__<server>__ in front of each tool name. The second column counts that variant.

The estimate column covers all 25 509 alive servers. 8 823 alive servers publish a schema the endpoint refuses; they have the estimate only.

Servers per cost band

How many of the counted servers fall into each band of exact tokens.

Exact against the estimate

The estimate serializes each tool as compact JSON and counts it with the o200k_base tokenizer. Claude renders the schemas in its own format, so the exact count is higher. The ratio is steady across servers.

The heaviest tool sets

Claude's context window holds 200 000 tokens. The last column says how much of it one server takes before the first prompt. 6 alive servers do not fit into the window at all.

All servers by token cost → · What a server costs, the short answer →

4. Auth

A server that answers 401 or 403 is reachable, but it does not talk to a client without credentials. The probe records the WWW-Authenticate scheme. Then it reads what the server publishes about getting in: the OAuth protected-resource metadata (RFC 9728) and the authorization server's metadata (RFC 8414). It never asks for a token.

How a client gets in

Four kinds. The sentence under each one says what it means for a client.

All servers that need auth → · How the kinds are read →

5. The lumps

The registry is not flat. Most hosts publish one server. A few hosts publish thousands of entries from one template. The table lists the hosts with at least 200 entries. “Distinct tool sets” is the number of different tool lists one host serves: a host with many entries and few tool sets publishes the same server under many names.

Duplicate tool sets

A hash of the sorted tool names and schema sizes marks servers that serve the same tools under different names.

Packages

The two package registries are the biggest hosts. A package is probed in a container with the network off and placeholder values for the environment variables the registry entry lists.

The 500 biggest hosts hold 22 961 of the 46 137 probed servers. The other 23 092 hosts hold the rest. All hosts →

6. Protocol versions

The version a server negotiates in initialize. The container run does not record the version of a package server, so package servers are the “unknown” row. A server that answers with a version the current SDK rejects is counted as “newer protocol” in section 1, not here.

133 servers answer with a version the SDK does not support yet. They are reachable and count as gated.

7. What the registry does not catch

The registry checks who may publish under a name. It does not check that the endpoint speaks MCP, or that it still exists. These servers are listed as active, and a client that adds one gets an error. Not MCP, not found, no DNS and paid are the big ones.

Packages that do not start

What changed since yesterday → · All servers that are down →

8. Method and limits

The probe connects to each server once a day with the official TypeScript SDK, sends initialize and tools/list, and sends no credentials. The exact token count comes from the Anthropic count_tokens endpoint: one request with the server's tools and a marker tool, one with the marker tool only, and the difference is the cost. A server is counted again only when its tool definitions change. A server that needs auth shows no tools, so its cost is not in the numbers. One probe from one network on one day can miss a server that is alive: a timeout or a rate limit on that day counts as down. Package servers run in a container with the network off, so a package that needs a real network or a real credential shows as gated, not as alive.

The full method → · The source →

9. The dataset

One row per probed server from the run of 2026-10-24, with the numbers the board shows for it. The numbers of a row are the numbers the JSON API gives for that server, /api/s/<name>.json, on that day. The file is a static file of this site: it is small, and a static file costs no Worker request on the free tier, so it is the cheaper of the two ways to publish it. A GitHub release would hold a file over 10 MB.

Download mcpcost-2026-10.json.gz

The document has the licence, the run date, the field list and the citation at the top. The rows are under rows. A missing value is null.

curl -sL https://mcpcost.ulehla.dev/data/mcpcost-2026-10.json.gz | gunzip | jq '.rows[] | select(.status == "ok") | .claude_tokens' | sort -n | tail -3

Fields

The same run as re-check evidence for the registry

The official registry discusses a periodic re-check of its entries (issue #1715). This file is this run in the row shape agreed there, so a registry job can merge it with evidence from other sources: entry, checked_at, check (endpoint for a remote server, package for an npm or PyPI server started in a container), state and evidence. The state is one of five: alive (initialize and tools/list succeeded), auth (reachable, needs credentials or payment), transient (the endpoint exists but did not serve this run: timeout, 429, 5xx, refused), dead (no DNS, TLS failure, 404, not MCP; a package that does not install), unknown (a package that needs values or a network to start). The evidence holds the probe status by cause, the target, the HTTP status, the auth scheme and kind, the protocol version, the tool count and the first line of the error.

Download mcpcost-evidence-2026-10.json.gz (1.6 MB, CC BY 4.0, the same licence and citation)

The newest run in the same shape is always at /data/mcpcost-evidence-latest.json.gz. It changes after each daily run. Its index, mcpcost-evidence-latest.index.json, gives the run date, the number of rows, the size and the SHA-256.

Licence and citation

The dataset is published under Creative Commons Attribution 4.0. You may copy it, change it and build on it, also for commercial work. Name the source. A link to this page is enough.

Ondřej Úlehla, The state of the MCP registry, October 2026. mcpcost, run of 2026-10-24. https://mcpcost.ulehla.dev/report/2026-10. CC BY 4.0.

The code of the probe and of this site is MIT. The daily summaries since 2026-10-03 are committed in the repository as data/latest/summary.json, and the summary API serves the latest one.