Skip to content

The state of vendors' MCP servers

Every day a program asks each Atlas vendor's own MCP server to start a session and list its tools. These are the counts from 2026-10-19: who answers, how they sign in, what the tools cost an agent's context and how well they are described.

numbers from 2026-10-19 · checks since 2026-10-19

Who answers4,271 servers checked on 2026-10-19
  • Answer without sign-in2,12350%
  • Ask for a sign-in2,11049%
  • Did not answer221%
  • Answered, but not MCP160%
How they sign in2,111
  • OAuth: the client signs the user in1,90990%
  • A key the user pastes1798%
  • Did not say231%
What the tools cost an agent's context2,123
  • Median server≈ 1,570 tokens
  • Heaviest one in ten≈ 9,254 tokens
  • Heaviest≈ 362,853 tokens

Every tool definition rides in the agent's context on every turn, before the agent has done anything.

The heaviest servers

How well the tools are described2,121
  • A1,32362%
  • B62630%
  • C1718%
  • D10%
  • F00%

A grade averages six checks of every tool: it says how well the tools are described, not whether they work.

Servers with at least one tool failing a check

  • Name outside the MCP spec's format0 · 0%
  • No description an agent can choose by17 · 1%
  • No object input schema0 · 0%
  • Parameters without a type40 · 2%
  • Parameters without a description947 · 45%
  • No read-only or destructive hint582 · 27%
Warning signs2,121
  • An instruction kept from the user0 · 0%
  • Invisible characters1 · 0%
  • Deletes, sends or pays without a hint133 · 6%
  • Marked read-only, but deletes or pays2 · 0%
  • Changes data on a server that answers without sign-in527 · 25%

A sign is a reason to read the tool, not a verdict. On the first day the rule for hidden instructions flagged twenty servers, and all twenty were honest guidance to the model; it now looks only for concealment.

Uptime over 30 days

Uptime is printed once a server has 7 daily checks; the first servers reach that on 2026-10-25.

Method

  1. Which servers: each Atlas vendor's own server in the official MCP registry, tied to the vendor by a verified namespace, by the GitHub organisation its homepage links, or by an endpoint on its domain. One server a vendor: the one its MCP page connects to.
  2. The check: once a day, initialize and then tools/list over streamable HTTP, without an account. A 401 is a sign-in, and its kind is read from the WWW-Authenticate header or the RFC 9728 protected-resource metadata.
  3. Context cost: the name, the description and the input schema of every tool, as JSON, at about four characters a token. It shows the order of magnitude; a tokenizer's count differs by model.
  4. Grade: six checks of every tool - a name in the spec's format, a description of at least 20 characters that is not the name, an object input schema, a type and a description on every parameter, a readOnlyHint or destructiveHint. The score is the mean, over the checks, of the share of tools that pass: A from 90, B from 75, C from 60, D from 40.
  5. Warning signs are found by rule in the names and descriptions, parameters included. A server behind a sign-in lists no tools to a stranger, so it has no grade and no signs here.
  6. No language model is used anywhere in this report.

Data

The same numbers as JSON, refreshed daily. Free to cite with a link to this page. Open the JSON

Every server, with its own measurements →