Runwire is an agent-native API workspace where developers and WebMCP agents build, execute, inspect, repair, and verify multi-step API workflows together.
The problem
Traditional API clients make dependent workflows tedious. Developers repeatedly copy identifiers between requests, diagnose failures, change configuration, rerun steps, and manually preserve evidence. Generic browser agents are unreliable here because they infer actions from interface controls and often operate without a shared, inspectable execution record.
The solution
Runwire exposes 22 typed, page-defined WebMCP tools representing real application capabilities. An agent can inspect requests, select flows, execute individual steps, forward extracted response values, diagnose errors, apply bounded repairs, run controlled reliability tests, and inspect run history without relying on browser-control clicks.
The demonstration uses a Checkout recovery workflow:
Create customer → Create order → Get order → Delete order
Runwire automatically extracts and forwards customerId and orderId. The first execution intentionally fails with 400 MISSING_IDEMPOTENCY_KEY. The agent reads the actual request and response evidence, applies the available idempotency repair, and reruns each request successfully with 201 → 201 → 200 → 204.
Every tool call, request, response, status, duration, extraction, failure, and repair remains visible in the interface. Failed and passing runs are preserved together, giving people a trustworthy audit trail of what the agent changed and why.
Why WebMCP
WebMCP lets Runwire expose semantic application actions instead of forcing agents to infer intent from pixels or DOM controls. The person and agent operate the same workspace and see the same evidence. This makes dependent API setup, debugging, and verification faster while keeping the human in control.
Implementation
Runwire was created during the challenge using React, Next.js, Vinext, WebMCP's document.modelContext.registerTool, Drizzle ORM, and D1. Safety boundaries prevent private-network requests, hide authentication secrets from agents, and limit automated repairs and burst testing.
Testing
No account or credentials are required.
- Open the live application in ChatGPT's in-app browser.
- Ask the agent to select and inspect the Checkout recovery flow using only Runwire's WebMCP tools.
- Run the flow and inspect the
MISSING_IDEMPOTENCY_KEYfailure. - Ask the agent to apply the bounded idempotency repair.
- Run the four requests individually.
- Confirm the final 4/4 result and compare the failed and passing runs in history.
Built With
- chatgpt-sites
- cloudflare-d1
- drizzle-orm
- next.js
- react
- typescript
- vinext
- webmcp