BrowTool

Inspiration

I was inspired by OpenAI’s launch of Operator and wanted to better understand what makes it different. Since browser agents can be expensive to run, we started thinking about a lower-cost approach for repeated workflows, where we record a user’s actions once and turn them into a reusable MCP callable tool quickly and efficiently.

What it does

This platform records the action of the user to make a repeatable and parameterized MCP tool for any ai agent of your choosing. Allows seamless browser based context retrieval.

How we built it

We build this using Playwright codegen to record user actions. With this, we store the tool name, description, parameter and code to have tool accessible by our mcp server to use whenever needed by our agent

Challenges we ran into

We originally planned on building a custom browser extension that tracked DOM movement, then tried converting that output into Playwright code after. That pipeline was slow and messy, so we switched to Playwright Codegen, which records actions and generates clean Playwright scripts instantly.

We also hit a reliability issue where our agent would call an MCP tool and scrape HTML before the page finished loading. The fix was a systemic change on the MCP server side that forces a full page-load wait before any HTML extraction, which made the entire agent flow way more stable.

Accomplishments that we're proud of

Having a working MCP within the first 3 hours of the hackathon

How well our project works with minimal bugs during the development phase

The extremely fast creation of MCP tools ready within less than a minute for a agent to use

What we learned

We learned how important context is when building with AI. The best way to get strong context is by using MCP tools to search outside sources, so the model isn’t guessing and can actually ground its output.

What's next for MDR

Next, we’re launching beta testing and adding remote MCP tool hosting. We also want to explore more ways AI can generate higher-quality responses so users can trust it to complete real work end-to-end.