Skip to content

Choosing a connection

The question is not which protocol your system speaks. It is what your agent does.

Your agentConnect viaVerdict ceiling
Uses toolsMCP serverPARTIAL
Calls an APIOpenAPI documentPARTIAL
Clicks around a pageBrowserOBSERVATIONAL

Why the ceiling matters more than the protocol

Section titled “Why the ceiling matters more than the protocol”

RigorRun decides a verdict by reading your system after the agent has finished. A connection that can read state back can close that loop; one that cannot, cannot.

A browser is the case worth understanding before you start. A page saying “done” is a claim by the same system that would have to be wrong for it not to be done — so a browser cannot check its own work, and RigorRun’s types refuse to let a browser be its own verifier.

That is not a gap waiting to be closed. It is what a browser is.

You can attach both. RigorRun drives the browser so the agent works the way a person would, and reads a connected MCP server or HTTP API for the same system afterwards to decide what actually changed. The clicking is watched in the page; the verdict comes from records.

Most systems that look API-less have something that can be read.

Staging, a scratch instance, or a local copy. Mark the environment honestly when you connect it:

SafetyWhat RigorRun will do
ephemeralAnything. Made for testing, reset freely.
localAnything. Your machine, nobody else affected.
stagingAnything. Shared, but not the live system.
productionNothing that writes. Your agent still gets to try, and you see what it would have done.