What we found
133 of 151 products we drove have an MCP server, 127 of them their own.
Of those, 85 of 133 kept our agent out at the door. It got into 48 on its own.
The server answers once the agent has an account and a key. Getting those is where it stopped: a bot check on the sign-up form, a key that only a dashboard hands out, or an OAuth sign-in that waits for a person in a browser.
What an agent needs behind the server
An account it can open.
A sign-up an agent can finish, or an API call that opens one. Let agents sign up and get a key.
A sign-in it can finish.
The MCP spec's sign-in is OAuth.1 Let the agent register its client and finish the flow without a person clicking through it, or use a device code. MCP sign-in agents can finish.
A bot check that knows signed agents.
An agent that signs its requests can be let through where the sign-up page has a check. Let signed agents past your bot check.
Docs that say how.
Name the server's address, how to sign in and which tools need which plan, on one page an agent can read.
Check it yourself
Start from nothing.
A new email address, no account and no key. Add your MCP server to an agent and ask it to do one real task.
Count the times a person had to step in.
Each one is where an agent on its own stops. The first one is the one to fix.
Written by our agents from our drives and the vendors' own docs. How we write.
