Observed, not claimed
Every verdict comes with the evidence it was reached on. A check we can't resolve is shown as unresolved and left out of the score.
///// About
Why Backwatch exists
MCP sign-in fails quietly, where the server's owner can't see it. Backwatch checks it from outside.
- LicenseMIT
- Checks132, in four pillars
- Study1,000 registry servers, 3 Nov 2026
- TelemetryNone
- StatusPre-release, in calibration
///// The problem
Agents fail quietly
When someone adds an MCP server to ChatGPT, Claude or Cursor, the host signs in on their behalf, through a chain of steps the server publishes. When one step breaks, the host shows a generic error and the request never reaches the server's logs. Its dashboards stay green.
On 3 November 2026 we checked a random 1,000 servers from the official MCP Registry. Of those that use sign-in, about 1 in 9 break before a user ever sees a login page.
Backwatch checks from the side the user is on: outside, the way each host reaches the server.
Read the study→///// How we work
Four rules
No AI decides a score
AI may explain a finding. Scores come only from checks you can read and run yourself.
On the engineer's side
Every request declares the Backwatch User-Agent. It sends no credentials and no telemetry, and registers OAuth clients only when you ask, on your own server.
Open source
The CLI and its checks are MIT licensed. Read how a verdict is reached, or run it yourself.
///// Talk to us
Questions, bugs, a check we got wrong
Open an issue on GitHub. If Backwatch showed up in your logs, the bot page says what it requests and how to block it.
Open an issue→