CancelDesk

Inspiration

Cancelling a subscription should be easy, but it often isn't. Users have to dig through account settings, find the right membership page, read cancellation policies, decline retention offers, answer surveys, and click through multiple confirmation screens.

We built CancelDesk to explore what that experience could look like in a WebMCP-powered web. Instead of making an AI agent navigate a website the same way a human does, CancelDesk exposes structured tools that let agents understand what the site can do and take action directly.

What it does

CancelDesk is a subscription management dashboard built for both humans and AI agents.

A user can simply say:

"Cancel everything I haven't used in 60 days that costs more than $15 per month, but keep StreamBox. Show me before actually cancelling anything."

Through WebMCP, the agent can discover the user's subscriptions, identify the ones matching those conditions, calculate potential savings, protect StreamBox, check cancellation requirements, and prepare the cancellations.

CancelDesk then pauses when a meaningful human decision is required. For example, CloudPro presents a retention offer that the user must accept or decline. Final cancellations also require explicit authorization.

In our primary demo, a workflow that would require 24 manual UI steps can be handled with 2 WebMCP calls — 92% fewer interactions.

How we built it

CancelDesk is a responsive web application with 20 native WebMCP tools exposed through document.modelContext.

The WebMCP layer connects directly to the same domain services and application state used by the human interface. There is no hidden browser automation and no simulated DOM clicking behind WebMCP actions.

We also built several visual tools to make agent activity transparent:

• Agent Inspector shows real WebMCP calls, inputs, results, and execution time.
• Mission Control shows the progress and status of multi-subscription operations.
• Agent Flow visualizes the manual UI journey being bypassed by WebMCP.
• Human vs. WebMCP comparisons show the actual number of manual interactions versus semantic tool calls.

The production application is agent-agnostic and does not require an OpenAI, Anthropic, Google, or other AI API key. The intelligence comes from the agent using the site; CancelDesk provides the capabilities through WebMCP.

Challenges we ran into

One of the biggest challenges was designing WebMCP actions that were powerful without giving an agent unlimited authority.

Cancelling a subscription is consequential, so we separated preparation from authorization. An agent can research policies, calculate savings, prepare cancellations, and surface retention offers, but preparation alone cannot complete a cancellation.

We also had to make multi-subscription operations safe against forged confirmation tokens, replayed authorizations, protected subscriptions, unresolved retention decisions, and partial failures.

Another challenge was making the benefit of WebMCP visible. Because a successful WebMCP interaction actually removes UI navigation, we created Agent Flow and interaction comparisons so users can see the manual steps that were avoided.

Accomplishments that we're proud of

We're especially proud that CancelDesk doesn't just demonstrate WebMCP with a single simple action. It supports a complete multi-step, multi-provider workflow while preserving meaningful human control.

CancelDesk exposes 20 native WebMCP tools, and our test suite verifies native discovery and execution across all of them.

Our primary workflow reduces 24 manual UI steps to 2 WebMCP calls, while still stopping when human judgment or final authorization is required.

Most importantly, the WebMCP implementation and human interface share the same underlying application state. When an agent takes action through WebMCP, the user can immediately see what happened in the actual application.

What we learned

WebMCP changes the relationship between websites and AI agents.

The biggest lesson for us was that the goal shouldn't simply be to make agents better at clicking websites. The more interesting opportunity is to let websites expose what they can do in a structured, semantic way.

We also learned that reducing interaction doesn't have to mean reducing user control.

WebMCP can eliminate unnecessary interaction without eliminating necessary consent.

What's next for CancelDesk

CancelDesk currently uses simulated subscription providers to demonstrate the complete WebMCP experience safely and deterministically.

The next step would be connecting the same architecture to real subscription providers and billing platforms as they expose secure APIs and agent-accessible capabilities.

The larger vision is a web where users don't have to learn every company's interface just to accomplish a task. Humans should be able to describe what they want, agents should be able to understand what websites can do, and websites should remain in control of how those actions are safely performed.

Built With

• chrome
• document.modelcontext
• html
• javascript
• json
• next.js
• playwright
• react
• tailwind
• typescript
• vitest
• webmcp
• zod
• zustand