Bragproof

The state of MCP servers

What happened when 1,346 MCP servers from npm were each installed on a clean machine and started with no configuration. Live numbers: they change as reports come in.

65%start with no setup (878 of 1,346)
213contact something when installed and started; 136 of them the same host
168have tool text with commands in capitals or other unusual wording
8have tool text about the user, other tools or private files

What happens when you start one

Starts with no setup
Installed, started with no arguments, no environment and no input, and listed its tools.
878 65%
Needs setup before it starts
It exited or refused to answer until it is given configuration, such as a key or an argument.
266 20%
Did not start
It exited or failed before answering, and its output does not say what it needs.
134 10%
Needs another program or service running
It tried to run or reach a program or service that is not in the runner's sandbox, which holds Node and little else.
22 2%
Did not install
npm could not install it the way the runner installs: from the registry only, without install scripts.
20 1%
Not for this machine
The package says it does not support the runner's machine: Linux on x64 with Node 24.
16 1%
Needs a build step at install
It needs code that is built by install scripts, which this runner does not run.
6 0%
Run did not finish
The run produced no report: it did not answer within the time allowed, or what it produced could not be kept.
4 0%

Browse the ones that start · the ones that need setup

Do they answer when called?

For 578 servers the runner found a tool it was willing to call blind: one that takes no input and reads like looking something up. Called once with nothing, 374 (65%) returned a result not marked as an error from at least one such tool. The rest returned an error, asked for setup, or did not reply in time.

What they contact when installed and started

gateway.pipeworx.io136 servers
github.com13 servers
release-assets.githubusercontent.com13 serversa release file downloaded from GitHub
api.github.com4 servers
api.ainative.studio3 servers
eu.i.posthog.com2 serversusually analytics or error reporting
huggingface.co2 servers
us.aws.cdn.hf.co2 servers
us.i.posthog.com2 serversusually analytics or error reporting
169.254.169.254:801 serverthe cloud metadata address; cloud SDKs probe it when looking for credentials
adobe-premiere-mcp-telemetry.hetkp8044.workers.dev1 serverusually analytics or error reporting
agent.tollylabs.com1 server
agent402.tools1 server
agentum.asimplek.cloud1 server
answerrail.com1 server

Hosts looked up or connected to while a server was installed and started with no input, watched from outside its sandbox. The npm registry is left out. Which servers

Settings they ask for most

GOOGLE_CLIENT_ID4 servers
BITBUCKET_TOKEN2 servers
BITBUCKET_USERNAME2 servers
CONFLUENCE_API_BASE_PATH2 servers
CONFLUENCE_API_TOKEN2 servers
CONFLUENCE_HOST2 servers
DD_API_KEY2 servers
DEEPSEEK_API_KEY2 servers
EMAIL_USER2 servers
IMAP_PASSWORD2 servers
IMAP_USER2 servers
JIRA_API_TOKEN2 servers
LINEAR_API_KEY2 servers
MCP_EMAIL_ADDRESS2 servers
MCP_EMAIL_PASSWORD2 servers

Names found in the error output of servers that would not start.

How this is measured

Every report links to the public run that made it and can be checked without trusting this site. How · Source · These numbers as JSON