GoodFolder

• 

  WebMCP agent proposes a reviewable GoodFolder change.

• 

  GoodFolder returns structured workspace context to the agent.

• 

  GoodFolder — accountable agent work, right where your files live.

• 

  Meet the GoodFolder mascot.

Inspiration

I built GoodFolder because I kept coming back to the same problem: once an agent touches a folder, it should be possible to understand what happened afterwards.

Files alone don't tell that story. GoodFolder keeps a readable history of changes and gives you a way back when something needs to be corrected. With agents involved, that history matters even more.

WebMCP felt like the missing piece for GoodFolder. A lot of the work inside a folder needs visual judgment: reading a document in context, checking a spreadsheet, looking at an image, or noticing that something simply doesn't look right. The agent can investigate and prepare work, while the person sees the same workspace and decides what should actually happen.

What it does

Open the challenge workspace in ChatGPT Desktop’s in-app browser. There’s no signup or setup. It opens straight into a seeded workspace with folders, files, saves, a timeline, and open change proposals.

Try asking the agent to list the folders, inspect the Q3 Report and its figures, trace a save, or prepare a correction. It can read the workspace and create a proposal or comment for review.

It can't quietly save a change, approve a proposal, restore history, remove files, or change access. A person still makes those calls.

How we built it

GoodFolder is a Next.js dashboard deployed on Cloudflare Pages. I added WebMCP Site tools through document.modelContext.registerTool, which lets ChatGPT work with the things GoodFolder already understands: folders, files, document structure, tables, history, and proposals.

The point wasn't to give an agent a faster way to click around a UI. I wanted it to understand what it was looking at and leave a clear trail when it suggests work.

The public repository includes the challenge work and documents the approval boundary.

Challenges and lessons

The difficult part was deciding where to stop the agent.

Reading a folder, tracing a save, and drafting a proposal are helpful. Letting an agent make lasting changes without review would defeat the reason GoodFolder exists. So I kept the agent close to the work, while making proposals the handoff point for anything that deserves a human decision.

WebMCP fit that model well. The page can give the agent real product actions without handing it the keys.

What’s next

I want to make the Dashboard and WebMCP experience feel tighter together: faster, clearer handoffs between what a person sees in GoodFolder and what an agent can inspect or suggest.

I’m also working on support for more file types and stronger previews in the Dashboard, so more of the work can be reviewed directly inside GoodFolder.

Built With

• cloudflare-pages
• next.js
• typescript
• webmcp