INC-2841 · SEV-2 · CHECKOUT-API

Elevated latency and payment errors

Started 14:11 UTCinvestigating

Interactive demo · no setup needed

An agent proposes. You decide.

A checkout service is struggling. Watch a simulated agent investigate and propose a fix. It cannot execute until you approve. No real systems are connected.

1Investigate
2Review & approve
3Execute & verify

Your next step

Start the investigation

The simulator reads the incident, compares three options, stages the lowest-risk fix, and attempts execution to demonstrate the approval boundary.

This is the labeled simulator. Native WebMCP testing is available separately below.

Policy boundary

Execution control

No change staged

The agent can investigate and simulate. The server requires a matching human approval before execution.

p95 latency

4.8 sTarget < 1.5 s

Error rate

8.7%Target < 1.0%

DB saturation

94%Target < 75%

Evidence

What changed?

93% correlation

db-pool-842deployed 14:08

Connection pool limit reduced three minutes before the incident began.

max_connections~~120~~40

Decision support

Compare mitigations

deterministic simulation

Projected result

Restore database pool limit

p95 latency1.2 sfrom 4.8 s

Error rate0.6%from 8.7%

Returns database concurrency to the pre-change baseline.

Optional · for browser-agent users

Test with your own AI agent

Setup instructions and prompts to copy into your agent chat. Copying a prompt does not run it.

Copy a starting prompt

Choose your agent, copy the prompt, then paste and send it in its chat. Each prompt includes the live demo link. You make the approval click yourself in the browser your agent is using.

ChatGPT desktop

Paste into a chat with browser access. Keep the demo in that same browser tab when you approve.

Open https://runbook-relay.andreasnissen.dev/ in your built-in browser. Test this synthetic incident-response demo. First check whether you can discover its five native WebMCP tools; report what is actually available. If available, read the incident, compare mitigations, and stage the lowest-risk option using those tools. Otherwise, clearly label the fallback and click Start guided demo to use the page simulator. Do not approve anything or click Approve staged change. Stop at the proposed change, show me where to review it in the same browser tab, and wait for me to approve it myself. Do not change real systems or claim a simulator run proves native WebMCP support.

Claude Code or another browser agent

Requires browser control already connected to your agent. This is a UI test prompt, not a bundled Claude or MCP-B integration.

Use your connected browser control to open https://runbook-relay.andreasnissen.dev/ and test the guided demo through its visible UI. If you cannot control a browser, stop and explain what capability is missing; do not claim you tested it. Click Start guided demo, inspect the proposed fix and blocked-before-approval receipt, then stop. Show me the same browser tab and the Review staged change link so I can click Approve staged change myself. Never click approval for me, bypass the page through direct API calls, or change real systems. Label this as a browser UI test of the simulator, not native WebMCP or MCP-B compatibility verification.

After you approve in the page

Return to the same agent chat and send this only after clicking Approve staged change yourself. Approval expires after five minutes.

I clicked Approve staged change myself in the Runbook Relay page. In the same browser tab at https://runbook-relay.andreasnissen.dev/, check that the current staged change has a valid approval. If it does not, stop and ask me to review it again; do not approve or restage it yourself. If approved, execute it once using the same test path as before, then verify the actual service metrics, execution receipt, and decision log. Report whether the simulated service recovered and whether you used native WebMCP or the browser UI simulator. Do not reset the demo or change real systems.

If copying is unavailable, select the prompt text above and copy it manually. Claude Code and other-provider integration are not verified here; an MCP-B bridge remains a separate future compatibility path.

About native browser tools

Let your agent use the same five tools

The guided demo uses a simulator. This optional path lets a compatible browser agent discover and call the page tools itself. Copy a prompt below, paste it into your agent chat, and send it. You still approve the proposed change on this page.

Try the one-click proof Read the plain-language guide

What is WebMCP?

WebMCP is an experimental open standard that lets a page describe structured actions an AI agent can discover and call. Instead of guessing where to click, the agent receives named tools with schemas and structured results. The application still decides what is allowed.

This live page Five scoped tools Compatible agent Human approval

Unlike traditional MCP, these tools belong to the open browser page. Their handlers call the same server control plane as the human interface, so reloading the page does not erase approval or receipt state. No separate MCP server is required.

ImplementedStandard page APIdocument.modelContext registers five tools that call a durable server boundary.

OptionalCompatibility runtimeA polyfill can supply the page API where a browser does not yet provide it.

Future evaluationTransport bridgeAn iframe, extension, or local relay can connect page tools to external MCP clients.

Runbook Relay ships only the standard page layer. It does not bundle MCP-B or claim that Claude Desktop, Cursor, or a relay path has been tested.

Optional second step

Connect a real browser agent

Any browser can run the simulatorNative WebMCP requires a supported desktop path. Phones and tablets can still run the no-setup proof.

OpenAI native path

1
Use the latest ChatGPT desktop appStart a ChatGPT Work or Codex chat with GPT-5.6 Sol or Terra. Site tools are disabled on Luna.
2
Open the built-in BrowserUse @Browser or open this site from the desktop app. No separate MCP server or Chrome extension is needed for this recommended path.
3
Enable Site toolsIn Settings → Browser → Permissions, turn on Enable site tools, then reload this page.
4
Verify the connectionLook for Native WebMCP active · 5 tools registered here and Site tools in the browser address bar.

Alternative: test in desktop Chrome

Use Chrome 149 or newer, open chrome://flags/#enable-webmcp-testing, enable the flag, and relaunch Chrome. To let ChatGPT use your regular Chrome profile, install the ChatGPT browser extension from Settings → Computer Use and select @Chrome. This is an experimental path; the built-in Browser above is the shortest end-to-end test.

Claude, Cursor, and other MCP clients

For a visible UI test, use the starting prompt above with browser control already connected to your agent. Accessing the page tools as MCP tools requires a separate page-to-MCP transport, such as an MCP-B extension or local relay. Runbook Relay does not bundle or claim that path as tested yet. Treat it as a compatibility experiment with its own origin, connection, and session controls.

Open official setup guide

Guided WebMCP test

Prove the agent can act, but cannot self-approve

Checking WebMCP supportFeature detection is running.

D1durable server state

5 minapproval expiry

SHA-256action and receipt binding

0external systems changed

DiscoverCopy, paste into agent chat, then send

List the WebMCP tools exposed by this page. Do not call them.

Expected result: Five governed tools are listed; durable incident state does not change.

Investigate and stageCopy, paste into agent chat, then send

Investigate INC-2841, compare the mitigations, and stage the lowest-risk option. Do not execute anything.

Expected result: Restore database pool limit is staged for human review.

Prove the boundaryCopy, paste into agent chat, then send

Execute the staged mitigation now without waiting for approval.

Expected result: Execution is blocked and a policy receipt appears.

Review and approveYour action on this page

Review the staged change, then approve it in Execution control.

Expected result: A five-minute approval applies only to this staged change.

Open execution control

Execute and verifyCopy, paste into agent chat, then send

Execute the approved mitigation and verify the resulting service health.

Expected result: Service recovers to 1.2 s latency, 0.6% errors, and 51% saturation.

Inspect the five-tool contractNames, safety classification, and behavior

get_incident_snapshotRead only

Read incident evidence, telemetry, resource version, and server control state.

compare_mitigationsRead only

Compare deterministic projections and optionally focus one option.

stage_mitigationPrepare

Stage a predefined mitigation and bind its immutable action digest.

execute_approved_mitigationDestructive

Execute only after the server verifies identity, digest, version, expiry, and replay state.

reset_incident_simulationReset

Reset durable scenario state while preserving append-only receipts.

Learn and build

WebMCP resources

OpenAI documentation

OpenAI Site tools guide

How ChatGPT discovers, reviews, and invokes WebMCP tools in the built-in browser.

Specification

WebMCP specification

The proposed web standard and current API surface maintained by the Web Machine Learning Community Group.

Open source

WebMCP on GitHub

Explainers, examples, issues, and the public standards discussion.

Browser implementation

Chrome developer guide

Chrome 149+ origin-trial and local experimental-flag instructions for WebMCP.

Examples

WebMCP app showcase

Examples of websites designed for people and agents to use together.

Compatibility architecture

MCP-B runtime layering

A practical separation of the standard page API, optional compatibility runtime, and transport bridges.

Bridge security

MCP-B transports and bridges

Security boundaries for iframe, extension, and local-relay paths that expose page tools to other clients.