Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for RELAY — Proof Before Power

RELAY — Proof Before Power

Human-governed WebMCP authority for incident response.

Inspiration

AI agents are increasingly capable of investigating incidents and proposing production actions.

But sensitive execution creates a governance problem: an ordinary approval workflow can leave the underlying agent capability continuously available, with only a permission check standing between investigation and production change.

RELAY explores a different model:

What if the human decision changed what the agent was actually capable of doing on the web?

What it does

RELAY gives an agent nine persistent WebMCP capabilities for incident investigation, evidence gathering, deployment comparison, hypothesis building, recovery preparation, and verification.

The sensitive production execution capability does not exist.

The agent can investigate the incident and stage one bounded recovery proposal, but it still cannot execute production recovery.

Only the visible human action:

AUTHORIZE THIS ACTION

changes the live WebMCP capability surface.

At that moment:

09 → 10

One exact temporary capability is registered:

execute_authorized_recovery

The agent can use it once to carry out the approved recovery.

In the demo:

• deployment #4821 becomes #4819
• error rate falls from 31% to 0.8%
• recovery is verified
• execution count becomes 1

Immediately after execution, the temporary capability is removed:

10 → 09

A replay cannot execute another recovery.

Why WebMCP

RELAY is not simply an API call hidden behind an approval button.

The important state is the actual browser-visible WebMCP capability surface.

Before authorization, the production execution tool genuinely does not exist.

After the human decision, one exact capability appears.

After its single use, it disappears.

WebMCP makes that authority native to the interaction between the web application and the agent.

How it improves the user experience

The human does not need to manually perform the entire incident response.

The agent can investigate, assemble evidence, compare deployments, form a hypothesis, and prepare a bounded recovery proposal.

At the same time, the agent does not receive standing production authority.

The consequential decision remains human.

The resulting flow is:

investigation → human judgment → bounded execution

rather than forcing users to choose between full manual operation and persistent agent authority.

What humans and agents can do together

The agent handles the evidence-heavy investigative work.

The human decides whether production authority should come into existence.

After authorization, the agent completes the exact approved recovery without requiring the human to manually execute the production action.

Then the authority disappears.

RELAY expresses human governance as a change in agent capability, not merely as an approval flag.

How we built it

RELAY uses real document.modelContext WebMCP registration.

Nine persistent tools support investigation and recovery preparation.

The state-changing recovery execution capability is intentionally absent.

A human-only UI authorization dynamically registers execute_authorized_recovery for the exact approved recovery.

After its single execution, that capability is removed.

Final live QA verified the registry lifecycle directly:

09 → 10 → 09

The execution capability was absent initially, appeared only after human authorization, executed once, and was genuinely removed afterward.

What's next

RELAY is a focused demonstration of ephemeral, human-governed agent authority.

The same capability pattern could be explored for other consequential web actions where investigation can be delegated broadly but execution should be created narrowly, explicitly, and temporarily.

Built With

• css
• html
• next.js
• react
• typescript
• webmcp