← cd ../projects

AI-Powered Web Automation Agent

A FastAPI service that turns plain-language instructions into multi-step browser automation — a Claude agent driving Playwright over MCP, deployed on Fly.io.

PythonPlaywrightFastAPIMCPDocker

Overview

An autonomous browser agent that drives real websites to complete multi-step tasks — data extraction, form submission — without hand-written scripts per task.

How it works

  • A Claude agent over MCP decides the next action from the current page state rather than following a fixed script, driving a real browser through Playwright.
  • State recovery and retry logic so a flaky step or an unexpected page doesn't sink the whole run — the agent re-observes and re-plans.
  • A FastAPI interface exposes task execution so other systems can trigger runs over HTTP, with no code changes required per task.
  • Containerized with Docker and deployed to Fly.io behind a token-authenticated endpoint.

What I learned

The hard part of "agentic" automation isn't the happy path — it's everything that goes wrong on step 7 of 10. Most of the engineering went into making failure observable and recoverable so long tasks actually finish reliably.